Before Target's First AI Chief Takes the Job
On Monday, August 24, Chandhu Nair is scheduled to arrive at Target as the retailer’s first chief AI officer. His public opening assignment is unusually plain. He plans to listen, learn, and build relationships across the company before deciding where to push.
The listening tour begins five days after Target reported $26.5 billion in quarterly sales and eleven days after it described a digital twin that had improved the availability of 63 fresh-food items in a small pilot. It also begins inside a company with nearly 2,000 stores, about 415,000 employees, a generative AI assistant announced for store teams, and a 2026 investment plan that adds store payroll alongside technology and AI.
Those numbers do not form one AI result. The 415,000 employees are not 415,000 users. The fresh-food pilot is not a chainwide return. Target’s roughly $2 billion of incremental 2026 investment is not an AI budget. Quarterly sales include merchandising, pricing, store hours, digital demand, capital projects, and a $994 million pre-tax tariff refund.
Nair’s appointment creates an opportunity to keep those categories apart before another executive dashboard joins them together. Target said his success should be judged through growth, guest experience, and the ability of teams to work more effectively, rather than the number of AI tools deployed. The company has not published the baseline, denominator, cost, or decision process behind those measures.
His first hundred days can supply that missing operating layer. Start with a map of workflows and human owners. Then connect each AI system to one customer measure, one worker measure, one financial measure, and a stop condition. The title creates a place to do the work; a portfolio charter would show whether it happened.
Monday starts with a listening tour
Target gave Nair thirteen days between announcement and start. He had spent about 25 years in data, technology, analytics, and AI roles, most recently as Lowe’s chief data, analytics, and AI officer. Purvi Shah is scheduled to become Target’s senior vice president of user experience on the same Monday.
Putting AI strategy and user experience on the same start date is revealing. A retail AI portfolio lives in technical systems and in moments a customer or employee can feel. Search and recommendations affect which products a guest finds. Forecasts decide whether the product reaches a shelf.
A store assistant changes how quickly an employee can answer a question. An inventory model can shift truck arrivals, receiving work, and replenishment priorities before a shopper enters the aisle.
Target’s appointment announcement gave Nair responsibility for advancing enterprise AI strategy and accelerating the use of AI across the business. It did not disclose his reporting line, budget, hiring plan, model authority, vendor commitments, or the systems that will move under his office. It also did not say whether product teams will own outcome targets or send them to the new executive.
An appointment announcement rarely fills those blanks. Their absence still changes the first task. A listening tour needs questions that expose ownership rather than a series of demonstrations.
For a store assistant, who owns an answer that conflicts with current policy? For a replenishment forecast, who can overrule the model when a regional event breaks the historical pattern? For conversational shopping, who distinguishes a new customer from a shopper who would have visited Target anyway? For employee tools, who measures correction work, training time, skipped recommendations, and the work transferred to another person?
CAIO appointments have spread quickly. An IBM survey of 2,000 chief executives, conducted across 33 geographies and 21 industries, reported that 76% of respondents’ organizations had a chief AI officer in 2026, up from 26% in the prior year’s survey. That vendor-sponsored, CEO-reported result is neither an audited census of corporate titles nor proof that a CAIO improves returns.
Academic work presented at AMCIS 2026 offers a narrower role hypothesis. Based on a systematic review, design work, and interviews with C-level executives, the researchers described the CAIO as a complement to the CIO, with a focus on strategic integration, resource allocation, and coordination across the organization. That is a model for asking about Target’s design, not evidence of the design Target chose.
Decision rights will determine whether Target added an operator or a shared label. A CIO can run reliable systems. A data leader can improve models. A product executive can ship an experience. A CHRO can manage training and role design. A store leader can protect service during a rush. Local incentives remain unchanged if the chief AI officer cannot resolve conflicts among them. Every team could report a successful launch while customer, employee, and finance measures point in different directions.
Listening is therefore a form of portfolio discovery. The output should be an ownership map that names the person who can scale, pause, or retire each system. Without that map, Monday becomes the first day of another coordination layer.
One title sits above several AI portfolios
Nair will not inherit one AI program. He will inherit several portfolios that reach different users, depend on different data, and produce different kinds of evidence.
Customers encounter one portfolio through search, recommendations, personalization, and conversational shopping. In June, Target said traffic arriving from AI-driven platforms had grown by more than 2,000% in the first quarter compared with a year earlier. The retailer had also made products available through experiences involving Google, Gemini, Microsoft Copilot, and ChatGPT.
Traffic is a useful acquisition measure. It leaves several commercial questions open. Target did not publish the starting traffic volume, the share of visitors who were new, conversion, basket size, returns, acquisition cost, margin, or the amount of demand that shifted from ordinary search and direct visits. A large percentage can emerge from a small base. A referral can also move an existing customer between channels without adding a sale.
During a shift, employees encounter another portfolio. Target announced Store Companion in 2024 as a generative AI assistant inside the handheld devices used by store teams. It could answer process questions, coach new employees, and help with tasks such as restarting a cash register or enrolling a customer in a loyalty program. The company planned to expand the tool from a pilot in roughly 400 stores to nearly 2,000 stores by August 2024.
Farther upstream sits Proxima, Target’s digital twin for middle-mile logistics. Its models represent facilities, inventory, transportation, and operating constraints so teams can test a change before making it in the physical network. Unlike most of Target’s public AI claims, Proxima has a disclosed pilot result and a bounded sample.
Headquarters work is harder to see. Target’s strategy materials refer to AI in planning, productivity, and operations, but the public record does not enumerate internal systems, their employee populations, or realized labor effects. Some may automate a task. Others may produce a recommendation, summarize material, or improve a forecast. Those forms of assistance place different demands on review and accountability.
Put the known portfolio into one provisional map:
| Portfolio | Public example | Closest disclosed measure | Missing decision evidence |
|---|---|---|---|
| Customer acquisition | AI-platform referrals and conversational shopping | More than 2,000% year-over-year traffic growth from AI-driven platforms | Starting volume, incrementality, conversion, margin, repeat use |
| Store work | Store Companion on employee handhelds | Planned expansion from about 400 stores to nearly 2,000 | Eligible and active users, answer accuracy, resolution, time, rework, training, employee response |
| Supply chain | Proxima digital twin | 2.5% on-shelf availability lift across 63 fresh items; about 98% flow simulation accuracy for one facility | Sustained lift, control group, labor effect, waste, margin, network transferability |
| Enterprise work | Planning and productivity applications | No consolidated public outcome measure | Workflow inventory, cost, hours, errors, overrides, role changes |
This map is incomplete on purpose. Target may measure these systems internally; retailers often keep operating data private because it reveals process advantages. Store Companion may be producing useful answers every day. Conversational shopping may be winning profitable customers. Those results simply do not appear in the public record.
Secrecy and internal control can coexist. The new office still needs a common private ledger. Without it, portfolio review favors whatever team arrives with the clearest demonstration or the largest top-line percentage. A 2,000% traffic claim can dominate a quiet improvement in employee search time even if the latter reaches more decisions. A chainwide rollout can look larger than a 63-item pilot despite the pilot having better outcome evidence.
Each portfolio also has a different failure radius. A weak recommendation may lower conversion. A wrong store-policy answer can create a customer dispute or compliance problem. A bad logistics simulation can move inventory and labor across facilities. An internal summary can distort a manager’s decision without appearing in a customer metric.
One executive can compare those risks only after the company standardizes the questions. Who receives the output? What decision follows? How often does a person override it? Which cost and correction work sit outside the product team’s calculation? What happens to the workflow when the system is unavailable?
Every portfolio can use the same evidence contract without sharing a model.
Store tools still lack a worker denominator
Store Companion reached its scale promise before it reached a public outcome denominator. Target’s 2024 announcement expected hundreds of thousands of employees across nearly 2,000 stores to gain access through their handheld devices. The tool had been tested in about 400 stores, and Target said early feedback was positive.
Access is the top of an adoption funnel. It does not reveal how many employees were eligible, trained, weekly active, able to resolve a question, or willing to use the answer without asking a colleague. Target has not published those 2026 denominators. Nor has it published answer accuracy by question type, escalation rates, time saved, time spent checking responses, customer outcomes, or changes in new-hire ramp time.
Seasonality makes denominator choice consequential. Target’s latest annual filing reported approximately 415,000 full-time, part-time, and seasonal employees as of January 31, 2026. Staffing rises during peak periods. Roles differ across stores, distribution facilities, headquarters, and other operations. Shipt also uses independent contractors, who do not belong in the employee count.
Using 415,000 as the denominator for a store tool would include people who may never carry the relevant device. Using the number of stores would hide variations in staffing and use. Counting prompts would allow a small group of frequent users to look like broad adoption. The useful denominator is a cohort: employees in a defined role, at a defined set of stores, during a defined measurement period, who had access and enough training to choose the tool.
A numerator can mislead as easily. More questions can indicate adoption, unclear operating procedures, repeated failed answers, or new employees searching for basic information. A resolved task is stronger evidence than a prompt, though resolution alone says little about speed, accuracy, or customer response.
One of Target’s own examples makes the measurement problem tangible. An employee asks Store Companion how to restart a cash register. A useful answer should restore the lane faster than the old process, avoid a call to a supervisor, and keep the customer moving.
A plausible dashboard can still miss the cost if the answer is outdated, the employee repeats the steps, or another person later repairs the device. Following that single task from question to resolution would reveal more than a monthly prompt total.
Time saved also needs a destination. Fewer supervisor interruptions could create more coaching time, more customer help, a shorter backlog, or a lower labor requirement in the next schedule. Those outcomes distribute the benefit differently. A portfolio record should name which one occurred instead of treating every saved minute as interchangeable productivity.
Frontline AI research explains why a worker measure belongs beside the product measure. UKG surveyed 8,200 frontline employees in ten countries during 2025 and reported that 38% used AI at work, while 53% felt prepared to use it and 64% worried AI could replace jobs.
Those self-reported figures come from a multi-country vendor study. They cannot describe Target employees or prove that AI caused a change in stress or performance.
They do identify questions a rollout dashboard can miss. Does the tool remove search time or add verification work? Does it make a new employee less dependent on an experienced colleague? If so, does the experienced colleague gain time, lose mentoring responsibility, or receive different work? Does an incorrect answer create rework for the next shift? Can an employee challenge the output without being treated as resistant to adoption?
Measuring those questions can create its own worker risk. Individual prompt histories, overrides, and response times could become a performance file even when Target collected them to improve a tool. Cohort-level evaluation usually provides enough evidence for adoption and quality. Any individual-level review needs a stated purpose, limited access, a retention period, and notice to the employee before the data influences scheduling, discipline, or promotion.
Target invested additional money in store staffing and training during 2026. Its first-quarter update said more than 300,000 team members had completed guest-experience training, and it described changes to its MyDevice workflow. The company also said guest-service measures had reached three-year highs. Training, device changes, more payroll, merchandising, and local management all moved during the same period.
Together, those changes are good operational news and poor causal evidence. A customer can receive faster service because more employees were scheduled, a device reduced steps, training improved the interaction, or the store had fewer disruptions. Several changes can work together. A credible AI evaluation preserves the joint operating result while testing the tool’s contribution in comparable cohorts.
A store manager can offer a practical objection. Helpful answers should not sit in a laboratory while customers wait. If employees ask for the tool and service improves, a demand for perfect attribution can waste time.
A store need not wait for perfect attribution. Compare similar stores or workflows, record who had access and who used it, sample answer quality, count escalation and correction work, and watch a customer or task measure over a defined period. That evidence can distinguish a promising tool from an installed one while making employees’ extra review work visible.
Proxima gives the new office a measured pilot
Sixty-three fresh items give the new office its cleanest starting point. Target disclosed the boundary of the Proxima experiment along with its result.
Proxima simulates parts of Target’s middle-mile network, the movement between import warehouses, regional distribution centers, and stores. Target described a small test involving 63 fresh items. The pilot produced a 2.5% improvement in on-shelf availability, according to the company’s August 13 account.
Target also used Proxima before opening a receive center in Houston. The model represented the planned flow with about 98% accuracy, the company said, allowing the team to test processes and layouts before the facility began operating. These figures have clearer objects than a deployment count: a set of products, an availability measure, a facility, and a comparison between simulated and physical flow.
Both numbers arrive with visible boundaries. Sixty-three items represent a narrow slice of Target’s assortment. Fresh products face expiration, demand variability, handling requirements, and local conditions that differ from apparel or electronics. A 2.5% availability change does not disclose the starting level, test duration, control design, sales effect, waste, labor hours, or margin. Target’s 98% figure concerns simulation accuracy under its definition, rather than inventory accuracy or productivity.
Narrow scope helps here. Naming the boundary makes a small result useful and tells the next team what it must test.
For Proxima, the next questions can be operational. Did the availability lift persist after the pilot? Which intervention produced it? Did it require more handling or expedite work elsewhere? Did better flow reduce stockouts without raising spoilage? How did employee schedules, receiving congestion, trailer dwell time, and exception management change? Which parts of the model transferred to a second market?
A facility average can hide who absorbed the variation. Smoother inbound flow may remove overtime and congestion. A more aggressive schedule can instead push volatility to drivers, receiving teams, vendors, or a late shift. Capturing total hours and exceptions by location and handoff keeps an availability gain from borrowing labor outside the measured boundary.
Target said future versions could gain agentic capabilities that simulate scenarios and recommend actions. That language describes a direction, not a deployed autonomous system. The distinction should remain in the portfolio record. A digital twin that helps people test a layout has one decision boundary. An agent that proposes or executes network changes creates another.
Proxima also demonstrates why a chief AI officer cannot judge value from model performance alone. A simulation can match a facility flow and still fail to improve the business. It can improve inventory availability while pushing cost or instability into labor, transportation, or waste. Conversely, a modest model measure can produce a valuable operating result if it removes a costly bottleneck.
Product, supply-chain, finance, and affected employees each hold part of that account. Requiring those parts before the pilot becomes a scale story is work for the new office.
A scale decision should carry a transfer test. The original baseline stays fixed. The company names the new market, product category, facility constraints, employee cohort, expected mechanism, cost, and time window. Results from the new site are reported beside the pilot rather than merged into a cumulative success number. If the effect disappears, the team learns where the model depends on local conditions.
Measured transfer may look slower than declaring a chainwide transformation. It can outrun a broad rollout that later requires expensive correction. Proxima already gives Nair a concrete case for setting the portfolio’s evidence standard.
Growth and tariff refunds complicate attribution
The quarter waiting on Nair’s desk looks favorable at first glance. Net sales rose 5.3% from a year earlier to $26.5 billion. Comparable sales increased 3.8%, with store comparable sales up 2.7% and digital comparable sales up 8.7%. Non-merchandise sales, including advertising, marketplace, and membership revenue, grew more than 20%.
Inside the same earnings release sat a $994 million pre-tax benefit from tariff refunds. Target’s selling, general, and administrative expense rate rose by about 30 basis points. The company attributed pressure in part to higher compensation costs, including added field hours and incentive compensation, and to spending connected with capital projects. Capital expenditure reached $1.4 billion, up 27%.
An AI office enters this quarter with several favorable top-line trends and several rising inputs. That makes attribution harder. Digital growth can come from assortment, promotions, delivery, app changes, media, customer behavior, or AI-supported discovery. Store growth can reflect traffic, pricing, inventory, service, remodels, and local execution. A tariff refund affects profit without saying anything about model value.
Target’s March strategy adds another tempting total. The company planned roughly $2 billion in incremental 2026 investment: more than $1 billion of additional capital and about $1 billion of operating investment. It described hundreds of millions of dollars for store payroll and training, along with technology and AI work in discovery, personalization, and operations.
Calling that sum Target’s AI investment would be false. The plan includes stores, supply chain, payroll, training, remodels, technology, and other growth work. Even the technology portion is not automatically AI. Target has not publicly disclosed the AI share, current run cost, vendor and compute commitments, integration expense, or employee time spent building and supervising systems.
A financial ledger should therefore separate at least four layers:
- Direct AI cost, including models, compute, licenses, vendors, specialized staff, evaluation, and controls.
- Workflow cost, including integration, data work, training, employee use, review, correction, and management time.
- Operating investment that changes the same outcome, including store hours, incentives, process redesign, inventory, and capital projects.
- Observed result, with its time period and a comparison that supports or limits attribution.
Keeping those layers apart protects both sides of the case. An AI program cannot quietly claim a sales improvement created by added payroll or an easier comparison. A useful system also avoids dismissal merely because Target invested in stores at the same time. Joint changes can be managed without pretending one caused everything.
Outside Target, the value record remains mixed. PwC’s 2026 survey of 4,454 chief executives reported that 56% had seen neither higher revenue nor lower cost from AI over the prior year. Thirty percent reported revenue gains and 26% reported cost reductions.
These are CEO assessments across 95 countries, not audited project returns, and they say nothing about Target. They still show why a C-suite title and a favorable company quarter cannot substitute for workflow evidence.
Finance may resist a separate AI account if the technology is embedded across products. The objection has merit. Employees do not create value by charging every prompt to a central office, and a retailer can spend more time classifying tools than improving a workflow.
Materiality sets a workable boundary. Target does not need an invoice-level taxonomy for every feature. A portfolio view belongs on systems that require significant capital, reach a large employee or customer population, affect a consequential decision, or create a new risk. Small features can stay inside product economics. Large bets need a visible cost envelope and an accountable business owner.
Nair’s office can change the operating rhythm here. Quarterly review should ask which systems crossed the materiality threshold, which outcome moved, which other investments moved with it, and what evidence supports the next dollar. Revenue growth and deployment counts cannot answer those questions alone.
A first-100-days operating charter
Listening needs a deadline. By day 100, Nair can turn what he heard into a repeatable decision process without forcing every team to use the same model or metric.
Build the artifact with one row per material workflow rather than one row per model. A model can support several workflows with different owners and risks. A workflow can combine a model, rules, data, employee judgment, and ordinary software. The row follows the decision people are trying to improve.
| Charter field | Record at day 30 | Test by day 60 | Decide by day 100 |
|---|---|---|---|
| Workflow and human owner | Name the task, affected customer or employee, accountable operator, and executive with scale or stop authority | Confirm ownership during a live exception | Keep one accountable owner or resolve the overlap |
| User denominator | Count eligible, trained, and weekly active users by role and location | Compare use, non-use, and opt-out or override patterns | Explain adoption gaps before expansion |
| Baseline | Fix pre-use time, error, service, inventory, or conversion measure | Run a comparable cohort or staged rollout | Accept, revise, or reject the claimed effect |
| Human work | Record training, review, correction, escalation, mentoring, and work transferred to another role | Sample total workflow time and rework across shifts | Count net work change, not gross model time saved |
| Customer or inventory result | Choose one result close to the workflow | Watch quality, access, and unintended effects | Scale only with a sustained operating result |
| Cost | Set direct model, vendor, compute, integration, data, and payroll envelope | Compare observed cost with the original range | Fund, renegotiate, redesign, or stop |
| Risk and override | Name prohibited actions, escalation route, audit owner, and outage process | Test a bad answer, unusual demand event, and system outage | Close control gaps before wider authority |
| Evidence state | Label the row announced, accessible, adopted, measured, or attributed | Require source, date, cohort, and limitation | Publish the internal basis for scale, hold, or stop |
An evidence-state column prevents a familiar collapse. Store Companion was announced and planned for chainwide access. Proxima has a measured pilot. Conversational shopping has a traffic result. Those statements belong at different levels. None should inherit the strongest status from another program.
Day 30 is a portfolio census. Nair’s team would identify material workflows, their business owners, affected populations, vendors, current claims, and available baselines. Teams should be able to keep an initiative in the census even when the evidence is early. The purpose is visibility, not punishment for honest uncertainty.
The census should also name systems with no AI label that change the same workflow. Store payroll, MyDevice changes, training, routing rules, and inventory processes can alter the measure attached to Store Companion or Proxima. Recording those co-interventions makes a later result easier to interpret.
Day 60 is an exception test. A portfolio review based only on average use will miss the moments that create cost and trust problems. The store assistant should be tested against an outdated procedure, an ambiguous customer request, a device outage, and a question that requires escalation. Proxima should encounter a demand shock or operating constraint outside its training pattern. The team records who notices, who can override, and where the work goes next.
Worker evidence deserves its own review at this stage. The company should sample employees across tenure, role, shift, and store conditions. A tool can help a new employee while slowing an expert. It can make one shift faster by leaving exceptions for the next. It can reduce questions to supervisors while eliminating a useful mentoring exchange. None of those outcomes follows automatically from prompt volume.
Customer tests need similar variation. A shopping or service system should face common misspellings, regional products, assistive-device use, and queries in the languages Target expects it to support. Average conversion can rise while a smaller group receives worse answers. The charter needs a quality floor for those groups, followed by a named repair or escalation path when performance falls below it.
Day 100 is a capital and authority decision. Each material workflow receives one of three outcomes: scale, hold, or stop. Scale requires a defined cohort, a sustained operating measure, known cost, functioning override, and a worker account. Hold means the evidence is promising but incomplete, with a named next test. Stop means the effect did not survive, the cost is excessive, the control is weak, or the workflow no longer deserves attention.
A stop decision should count as portfolio work, not embarrassment. New executive offices often accumulate programs because every launch has a sponsor and few teams receive credit for ending one. An explicit stop category protects budget for the systems that survive testing.
The charter also changes the role of the central AI team. It does not take operational ownership from stores, supply chain, product, finance, or HR. It sets the evidence contract and arbitrates conflicts when one team’s local result moves cost or work into another function. The operator still owns the workflow.
Consider Store Companion. The store organization might own resolution and service. The product team might own answer quality and reliability. HR or learning teams might own training and employee feedback. Risk teams might define restricted answers. Finance might validate cost. The CAIO office would make sure one scale decision includes every account and that one executive can resolve a conflict.
Consider Proxima. Supply chain owns physical flow. The modeling team owns simulation quality. Store operations experiences the inventory result. Employees at facilities absorb changes in tasks and scheduling. Finance sees working capital, waste, transport, and labor cost. The charter does not flatten these measures into one score. It places them on one decision page.
Any charter can decay into paperwork. Teams can fill fields with polished language, select easy metrics, and postpone hard comparisons. Three safeguards keep it operational: link every row to dated source evidence, assign one person the authority to act, and set the next decision date before the review ends. Empty fields remain visible rather than becoming narrative footnotes.
Target does not need to publish the complete charter. Competitive details, security controls, employee data, and vendor pricing may require confidentiality. Public reporting can still improve. The company can distinguish announced access from active use, identify a pilot cohort, disclose an outcome with its scope, and explain whether cost and worker effects were measured. That is enough for customers, employees, and investors to understand what stage a claim has reached.
At closing time, the metric returns to the aisle
At closing time, fragmentation stops being an org-chart problem. Target’s systems cross product, store, supply chain, technology, finance, and workforce boundaries, yet their consequences meet in one aisle. A central executive can spot overlaps that a local team cannot. Common evaluation can reduce duplicate work, clarify risk, and move resources toward better programs.
From a store or product team, the same office can look like another reporting layer. It may centralize vocabulary while budgets and decisions stay inside existing functions. It may reward deployment counts because they are easy to collect, turning operating teams into data suppliers for a dashboard they do not use.
Nair’s own stated measure points away from that outcome. He said success should appear in growth, guest experience, and teams’ ability to work more effectively. Each phrase now needs a denominator and an owner.
Growth should distinguish incremental customers and margin from shifted traffic or unrelated quarterly gains. Guest experience should connect a workflow to availability, resolution, speed, quality, or trust. Team effectiveness should include eligible users, total task time, correction and escalation work, learning, schedule effects, and the distribution of gains across roles.
Monday’s listening tour will produce competing stories. One team will have a large traffic percentage. Another will have an employee tool distributed across stores. Proxima will have a narrow result with visible boundaries. Finance will have a broad investment plan and a quarter shaped by store spending and a tariff refund. Employees will have examples that never reach an executive slide.
The first hundred days should preserve those differences long enough to make a decision. If the office can show which workflows were merely announced, which were adopted, which changed an operating result, what they cost, and what happened to the work, it will have earned a place in Target’s operating decisions.
Closing time supplies the audit. Is the product on the shelf? Did the customer get a correct answer? Did the employee resolve the task without leaving hidden rework for the next shift? The evidence either makes that trip or disappears somewhere between the model and the aisle.
Related Reading: