Should an IT services firm build or buy an AI assistant for pay and HR questions?

Build one in-house assistant for pay, policy, leave and salary advances. It pays back about 19 months from kickoff and keeps payroll data inside the firm.

Situation

950 of 1,200 employees are consultants at client sites in 14 states. Every pay, leave and policy question goes to HR by email.

Complication

HR spends about 500 hours a month on 2,400 questions, answers take 1.5 days, and headcount grows 15% a year. Two vendors were pitching packaged assistants.

Resolution

Build one assistant on the firm's cloud, connected to Workday, ADP and one clean policy library. Release it in three phases, each gated on measured results.

Why

Most of what HR answers is already written down somewhere. Payslips, leave, policy and salary advances make up 8 in 10 questions, and each answer sits in Workday, ADP or a policy file. The plan targets 75% of those answered without HR, about 6 in 10 of all questions. The day 6 prototype got 92% of its answers right.

The biggest category is also the one nobody could sell us. A consultant's payslip question depends on their client, their rate, their timesheet and the state they worked in that week, and neither vendor could see that. Over three years from kickoff, building nets about $148,000. The service platform would have lost about $54,000.

The numbers hold when pushed: run-rate payback stays under 18 months whichever single assumption we stress. What worried us more was legal exposure on salary advances, and whether consultants would use the thing at all. Both now have an owner and a gate (Sections 7 and 8).

The payroll question that shaped the pilot

A representative case from the day 6 test set; wording and figures changed.

See the engineering handoff: the ticket, its tests and the scorecard, as the firm's team received them.

What changed our plan. Going in, I assumed the model would be the weak point. It wasn't. Of the 24 answers the prototype got wrong, 19 came down to data or policy: an old policy version, an ADP field we hadn't mapped, a rule HR had never written down. So we changed the order of work. Field mapping became a condition for launch, the policy clean-up moved in front of the build, and anything the data can't fully explain now goes to payroll rather than getting a best guess.

We agreed the decision criteria with the CFO on day 1, before looking at any technology.

Ten days, one output per day

HR spends about 500 hours a month on questions whose answers already sit in Workday, ADP or a policy file.

What drives the volume

Two things surprised us. The load isn't spread evenly: most payslip questions land in the three days after a pay run, exactly when the same two payroll people are approving advances. And many "payslip" questions are really tax questions. Consultants move between client sites, sometimes across a state line, and "why did my take-home change?" can only be answered from that person's own pay history.

The rest is housekeeping. Some policies exist in three versions (intranet, shared drive, old emails), so answers drift. Slow replies make more work too: one consultant in five asks the same question twice.

Four question types clear the bar. Timesheets wait for better data; visa letters stay with counsel.

Advances are worth more than policy or leave but go last: they move money, so the assistant earns trust on read-only questions first.

Building returns the most over three years, because only a build can read each consultant's own pay data.

We looked at four ways forward, including not using AI at all.

• A. Clean up the policies and publish a searchable FAQ. Cheap, and worth doing whatever else happens, so it's folded into phase 1.
• B. Switch on Workday's built-in assistant. It handles leave and policy well. It can't read ADP, though, and ADP's own assistant only covers pay inside ADP, so consultants would have two assistants that each know half the story. Advances need both.
• C. Buy an employee service platform on top of both systems. Good with policy documents; the payroll connectors were quoted as custom work.
• D. Build on the firm's own cloud, reading Workday and ADP directly.

Microsoft's Employee Self-Service agent was on the market, but it needs Microsoft 365 Copilot licences the firm didn't have, so we didn't score it.

$107,000 one-time pays back about 19 months from kickoff, 13 months at full run rate. The share of questions answered without HR is the number to watch.

HR's hourly cost moves payback the most, but it isn't really in doubt: finance gave us that number on day 1. What nobody can know yet is how many questions the assistant will close on its own. That is what the pilot is for, before the rest of the budget is spent.

For the CFO: the model values recovered HR capacity (280 hours a month, about 1.75 people). Cash shows up only in the hiring budget: the two HR generalists planned for next year's growth, about $170,000 a year, are not hired. That budget is larger than the capacity value, so the model is conservative. Faster answers and fewer pay errors are not counted.

No real employee identifier or pay figure ever reaches the model provider, and every action can be explained a year later.

How one question is answered

1. The employee signs in with SSO; the assistant can read only that person's records.
2. Names, IDs and amounts are swapped for masks ("EMP-4821", "AMT-07") before anything leaves the network.
3. The model drafts the answer from the question, the policy passage and the masks.
4. Inside the network, real figures are put back from ADP. Every pay figure comes from ADP's payroll APIs or from deterministic, tested code; the model only explains the result.
5. If no approved calculation covers the question, the assistant does not improvise: it routes the question to payroll with the records attached.

Policy rules HR can change, and when a person decides

Policies lived in documents and in HR's heads. The build turns the deciding rules into settings HR edits on a rules screen, with every change logged: who, when, old and new value, effective date. Each decision records the rule version it used. Anything outside the rules goes to a person; the assistant never approves an exception.

An audit trail for every action

Every action is logged with the request, records read, rule and policy versions, model and prompt version, decision and approver. HR and counsel can replay any decision from the last 12 months; older records follow the firm's retention policy.

What IT security asked

The biggest risks are legal and adoption, not the model. Each has an owner.

Three releases over 24 weeks, each opened only when the previous gate is met.

What the week 14 decision rests on. The pilot measures pay and policy directly. Leave is estimated from the prototype and advances from the written rules, so the committee releases the remaining $33,000 only if the measured results hit every target and projected run-rate payback is 18 months or less. Full scorecard in the engineering handoff.

Decision rights

Note: R responsible, A accountable, C consulted, I informed. Steering committee meets every two weeks.

KPIs

Three things the firm was considering should not be done now.

Timesheet questions can wait. Client approvals arrive late and only in a nightly export, so any answer would be stale. Fix the feed first, then look again; it's worth about $31,000 a year once the data is current.

Visa letters should never be automated. Each one has to match what was filed with immigration and needs counsel to sign it. A library of pre-filled templates saves more time, with almost none of the risk.

And we advised declining both vendor proposals as they stood. The platform costs more over three years and can't see consultant pay. Workday's assistant deserves another look in a year, sooner if it learns to read ADP.

Approve phase 1 this month and consultants use the assistant before quarter end.

Note: phase 1 covers engineering and pilot support to the week 14 decision, the security review, ADP API setup and policy clean-up.

Method, assumptions and lessons

A. Method

• Volume: 7,100 messages, masked in the firm's cloud, classified by a model; 400 checked by hand (94% agreement).
• Handling time: median of 41 timed requests, checked against HR estimates; the lower figure used.
• Feasibility: 300 masked questions on a day 6 prototype; 92% of payslip, policy and leave answers correct before clean-up; the 24 misses are broken down in Section 1.

B. Assumptions

C. From the lead advisor

On a comparable system I led, the hardest question was not which model to use but when a person must decide. HR policies lived in documents and in the HR head's memory, so we turned the deciding rules into settings HR could change themselves, with every change logged. Anything unusual went to a person: a medical expense claim over €2,000, for example, needed further proof, so it went to HR instead of being paid. We also kept a full record of every action, so a year later we could see what the system did and why. With pay and HR, any mistake can become a legal question, and that record is how you answer it.