The True Cost of HR Technology: A Sourceable TCO Model

Short answer

The subscription is one line in the cost of HR technology. A defensible total-cost model also includes implementation, data migration, integrations, security and privacy review, internal project time, training, recurring administration, AI or messaging usage, monitoring, incident handling, parallel systems and exit.

There is no credible universal multiplier that turns the contract price into “true cost.” The ratio depends on company size, current systems, jurisdictions, data quality, workflow complexity and the authority given to automation. Replace the multiplier with a source register: every cost has a unit, quantity, rate, timing, owner and evidence link.

This revision removes a fictional logistics-company CFO, an invented $187,000 cost autopsy and purported private vendor pricing behavior. It provides a model that a buyer can populate from quotes, payroll records, architecture and actual usage. Digidai did not perform the forensic review described in the previous version.

Define the boundary before calculating

Start with the decision the model must support. Examples include replacing an ATS, adding a scheduling tool, deploying AI sourcing, or consolidating recruiting and onboarding. State:

• business units, countries and worker populations;
• implementation start, production date and evaluation horizon;
• products being added and systems being retired;
• internal and external labor included;
• shared infrastructure allocated to the project;
• taxes, currency and discount treatment;
• residual commitments after termination.

Without a boundary, comparisons are easy to manipulate. One vendor quote may include onboarding and support while another excludes them. An incumbent may look free because its sunk implementation cost is omitted, or look expensive because all shared HR operations are assigned to it.

Keep cost and benefit separate. Total cost of ownership answers what resources the system consumes. A business case then compares those costs with incremental outcomes. Subtracting hoped-for savings inside the cost model hides whether the benefit occurred.

The cost equation

For a selected evaluation period:

TCO = contracted recurring charges
    + implementation and migration
    + integration and infrastructure
    + internal delivery labor
    + assurance and compliance
    + adoption and support
    + recurring operations
    + incident and correction cost
    + exit and transition cost
    - avoided costs from systems actually retired

Annualize one-time items only when the comparison calls for it, and show the assumed useful life. Do not silently spread a migration over five years when the contract can be cancelled after one, or charge a one-time replacement expense to every future year.

NIST’s AI Risk Management Framework explicitly asks users to examine expected and realized costs, including non-monetary costs from errors and system behavior. The current AI RMF Core also calls for third-party risk, human oversight, monitoring and safe decommissioning. NIST is a voluntary framework, not a pricing standard, but it supports treating assurance and exit as lifecycle work rather than procurement extras.

Cost register: what belongs in each bucket

Contracted recurring charges

Capture base subscription, priced seats, employees or candidates, active jobs, business units, storage, support tier, sandbox, analytics, messaging and AI consumption. Note minimum commitments, overages, renewal uplift, foreign-exchange exposure and taxes.

Public price pages reveal structure, not the final bill. Workable’s current pricing page describes product bundles, a trial and configurable AI-credit packages for larger volumes. That is a vendor offer and may change. It shows why “per month” is insufficient: the buyer needs the meter, included quantity, overage and feature dependency in the cost register.

Implementation and migration

Include solution design, configuration, data cleaning, field mapping, document transfer, validation, cutover and parallel run. Separate records that must migrate from records that can remain in a read-only archive. Retaining unnecessary candidate data increases work and risk.

Use a work breakdown rather than a percentage of license price. Count objects, integrations, countries, workflows and test cycles. Record who supplied the estimate and what acceptance criteria close the work.

Integration and infrastructure

Include connectors, middleware, identity, API gateways, data warehouse changes, logging, monitoring, environments, network controls and ongoing maintenance. A packaged connector still needs field mapping, error handling, permission review and regression testing.

For agents, add model tokens or requests, retrieval, search, browser or enrichment services, tool calls, retries, long-running state, evaluation runs and observability. Usage can grow with workflow loops even when user seats stay flat. Budget by completed action and by failed attempt, not prompts alone.

Internal delivery labor

Count time from HR operations, recruiters, hiring managers, IT, data, security, privacy, legal, finance, procurement, accessibility and change leads. Use loaded labor cost consistently. Distinguish new spend from capacity redirected away from other work, but show both so the implementation burden is visible.

Meeting time is not automatically waste. Decision, testing and exception design are part of the product. The cost problem is unowned work that repeats because evidence or acceptance criteria were missing.

Assurance and compliance

Consequential employment tools may require privacy impact assessment, security testing, accessibility review, job validation, bias assessment, notices, consultation, documentation and ongoing monitoring. The exact duty varies by use and jurisdiction.

The UK Information Commissioner’s Office reported almost 300 recommendations after audits of AI recruitment providers and published buyer questions about data protection, fairness, minimization and explanation. That is not a cost estimate. It is evidence that procurement and implementation need funded assurance work.

Australia’s privacy regulator similarly recommends due diligence, privacy impact assessment, human oversight and ongoing review in its guidance for commercial AI products. An annual security questionnaire alone does not cover those operating tasks.

Adoption and support

Include role-based training, manager communication, candidate support, updated procedures, office hours, help-desk load and productivity lost during transition. Completion of training is an activity. Test whether users can identify model limits, follow the review rule and recover from failure.

Support cost rises when the product’s exception paths are unclear. Track tickets by cause: configuration, data, integration, vendor defect, policy ambiguity or user misunderstanding. That evidence should change the next renewal and roadmap.

Recurring operations

Budget administrators, access reviews, configuration changes, data-quality work, model and prompt evaluation, release testing, policy updates, vendor reviews, reporting and deletion. AI features can change beneath a stable interface, so material-change notices and regression testing have real operating cost.

The owner must be named. “HR and IT” is not a resourcing plan. Assign capacity by recurring task and expected frequency, then replace estimates with observed hours after launch.

Incidents and correction

Track downtime, unauthorized messages, duplicate actions, wrong-stage moves, inaccurate inferences, discrimination or accessibility complaints, data exposure, payroll errors and candidate remediation. Use scenarios rather than pretending to know one expected number: plausible frequency, affected population, detection time, containment, correction and external support.

Public enforcement can show the type of downside without becoming a budget benchmark. The EEOC’s iTutorGroup settlement required $365,000 and injunctive relief to resolve allegations that application software automatically rejected older applicants. A settlement in one case is not the expected cost of another system. It demonstrates why automated rejection needs testing, monitoring and a remediation path.

Exit and transition

Price export, data mapping, archive access, deletion verification, connector removal, identity cleanup, parallel run, replacement implementation and post-termination support. Note which logs and decision evidence remain available after the contract ends.

Run an exit test before signing and before renewal. A statement that data is exportable is not evidence that the export is complete, documented and usable by the next system.

Build the worksheet from evidence

Use one row per cost driver:

Field	Purpose
Cost ID and category	Prevent duplicate or missing lines
Unit	Seat, employee, job, message, API call, hour, country or integration
Quantity and period	State volume and timing assumptions
Rate	Quote, contract, payroll or cloud price used
Scenario	Low, expected and high where uncertainty is material
Cash or internal capacity	Preserve budget and workload views
One-time or recurring	Support correct annualization
Source	Contract clause, quote, invoice, owner estimate or usage log
Owner	Person who can validate and update it
Confidence	Contracted, measured, estimated or unknown

Unknown is a valid value during diligence. It creates an action, not a zero. Ask the vendor, run a sample export, measure API calls in a pilot or allocate internal time explicitly.

Calculate benefits from completed outcomes

Avoid multiplying every automated minute by salary and calling the result savings. Time becomes financial value only if the organization redeploys capacity, avoids planned hiring, reduces paid external work, increases valuable output or improves a measured hiring outcome.

Use a benefit register parallel to cost:

• recruiter or coordinator capacity actually redeployed;
• agency, job-ad or contractor spend actually avoided;
• qualified hires started earlier, with the valuation method stated;
• fewer corrections, failed checks or duplicate tools;
• improved candidate completion and accepted offers;
• systems fully retired and contracts not renewed.

Keep attribution conservative. A hiring improvement may also reflect compensation, employer demand, a new recruiter or a different labor market. Use a baseline and comparison where practical, and show the period and population.

Unit economics reveal more than a single ROI percentage. Report cost per active job, completed application, interview, accepted offer and retained hire. For agents, add cost per successful authorized action and per corrected action. A cheap message generator can be expensive if it increases review and complaints.

Enterprise and SMB versions

An enterprise model should segment countries, business units, integrations and model uses. It needs central shared costs plus local implementation and compliance. Consolidation savings count only when old contracts, infrastructure and work are actually removed.

An SMB model can fit on one sheet, but it still needs subscription, setup, owner time, data, training, monitoring and exit. The main safeguard is scope: one or two systems, minimal customization, a manual fallback and no unreviewed consequential actions. A lower license price does not justify sending candidate data to an unapproved public tool.

In both cases, refresh estimates with actual invoices, usage and labor after the first operating period. TCO is a living model. Its value is the traceable decision, not the apparent precision of the total.

Correction and source scope

The September 13, 2026 revision withdraws the fictional MidWest Express case, named employees, invented invoices and the claim that a $42,000 subscription produced exactly $187,000 of spend. It also removes unsupported universal multipliers and private vendor-pricing assertions. The original file name, publication date and URL remain unchanged.

The Workable page is a current vendor offer, not an industry benchmark. NIST, ICO and OAIC materials describe risk and governance work, not prices. The EEOC settlement is a documented case, not an expected-loss estimate. Buyers must fill the model with their contracts, architecture, labor data and applicable legal advice.