Most design firms now have an AI line in the budget, and very few can say what AI cost on their ten largest projects last month. That gap was the subject of the talk Andriy Mulyar, Nomic's founder and CEO, gave to finance leaders from design firms at the Design Finance Officers Group fall meeting in Scottsdale on October 5. The room asked for practical material a CFO could hand to a controller, so that's what follows: where AI spend belongs on the books, how to bill it, how to measure what it returned, and what went wrong at firms that tried.
We sell AI on a usage basis, so we see this invoice from the seller's side every month, and the first question most CFOs ask us is how they'll control it and charge it back. Everything below applies no matter whose AI you buy. The field examples come from firms we work with, with the names taken out.
Most firms pay for AI, and few can show what it returned
Two surveys from this year frame the problem. The 47th annual Deltek Clarity A&E industry study (896 firms in the U.S. and Canada, released May 2026) found that 70% of A&E firms now use AI, up from 53% a year earlier, while only 38% report a measurable positive business impact. KPMG's Global AI Pulse for Q2 2026, a survey of 2,145 senior leaders at organizations with more than $50M in revenue, found that 49% had rephased an AI agent rollout when costs outran the expected value. Organizations with full visibility into their AI operating costs were five times as likely to report established ROI, 15% against 3%.
Read together, those numbers point at accounting. When AI spend lands in a bucket nobody can connect to the work, nobody can measure what it returned. If the spend isn't tied to a project, you can't tie it to a project outcome, and that's a finance problem finance can fix without much cost.
Four ideas to take back to your firm
The AI line will stay small compared with the labor it touches. What moves your P&L is where you book it, how you price the work, and what happens to the hours it frees up.
- Book AI usage the way you book job costs. It moves with project work, like printing, CADD time, or mileage, so put the project number on every run.
- The return depends on how the job is priced. Saved hours become margin on fixed-fee work and lost revenue on hourly work, unless repeatable scopes move to deliverable pricing. That's a pricing decision finance should be part of.
- Saved hours are capacity until you put them to work. They reach the P&L when teams take on backlog or more work at the same fee. If nobody refills them, utilization falls. The people stay, and the work that's been waiting gets done.
- You can't manage a return you can't see. The firms that show a return see spend by project, user, and workflow, which is why the project number matters more than the tool.
Engineers still review and stamp the work, so their review time stays on the cost side of every number that follows.
The AI bill changed shape
Your budgeting and AP process was built for seat licenses: you sign, you know the number, and it goes in overhead. Most AI products now bill some or all of their price by the unit of work, such as pages read, reviews run, and model usage. You learn the cost after the work unless someone estimates first, the volume and difficulty of the work drive it, and everyone who starts a task is making a spending decision.
Unit prices are small, so volume and mix drive the bill. Here is what common tasks cost on Nomic's default model:
| Task | Typical cost |
|---|---|
| Question on an indexed project | $0.10 to $0.15 |
| Code compliance check, one sheet | $0.15 to $0.40 |
| Drawing review | $0.25 to $1.00 per sheet |
| Submittal review, one spec section | $0.75 to $2.00 |
| Indexing a document, one time | $0.012 per page |
A 200-sheet set reviewed at $0.75 a sheet comes to about $150, roughly one billable engineer hour. The same person can spend $3 one month on lookups and $300 the next on two full-set reviews, and both are normal. Two people with identical seats produce very different bills, and the difference follows the project work they're doing, which is why the project is the right place to account for it. Other vendors price differently, but the shape is the same, and actual cost varies with file size, model choice, and output length. You can model your own mix with the AI usage estimator.
The AI line is small compared with the hours it takes off the work
These results come from live projects at firms we work with. The hours are what engineers and estimators reported, before their own review of the output, so treat them as a ceiling until your timesheets back them up.
| Workflow | By hand | With AI agents |
|---|---|---|
| Submittal review, 3,500 pages, at a global design firm | About four hours | Five minutes |
| IFC model health check, by a BIM lead at the same firm | Eight to 15 hours | About two hours |
| Scope extraction for a bid package, by a contractor's estimator | About 47 hours | 18 minutes, with 20 high-priority RFIs flagged |
Over a 12-week pilot at a global design firm, 42 engineers reported 669 hours saved and 19 critical errors caught before they turned into rework. AI cost on work like this runs from tens to a few hundred dollars per run, while the engineering hours around it are worth thousands. The gap between the two lines is two or three orders of magnitude, so the price of AI is rarely the number that matters.
The cheapest honest test of any AI vendor came from a contractor. It handed us a past job it had lost about $17,000 on, without telling us, and the review found the issues behind the loss. Pick a finished job where you already know the answer, and run every vendor you're considering against it.
Where AI spend belongs on your books
Firms solved this problem once already. When CADD stations showed up, they built charge rates, logged usage by job, and credited the overhead pool. The AASHTO Uniform Audit & Accounting Guide still asks, in item H.8 of its internal control questionnaire for consulting engineers, whether computer usage logs are maintained and coded by job or project. AI usage is that kind of cost, with better logs than CADD ever had. It can go to three places, and each one has rules.
- Overhead pool (indirect). Firm-wide tools, general productivity, training, pilots, and anything you can't tie to one project, recovered through the overhead rate on all work. It raises your rate, FAR Part 31 allowability applies, and no project manager owns the variance. Overhead is fine for general-purpose chat tools. It stops being fine when a meaningful share of spend is review work on a specific job, because then you're spreading one project's cost across every client and inflating your rate.
- Project cost (direct ODC). Runs done for one project and logged to its project number, billed at cost or at an established rate where the contract allows other direct costs. The same cost can't also sit in the pool, and some DOTs allow direct computer charges only as a pass-through with no profit.
- Capitalized implementation. Configuration, integration, and workflow build for a hosted AI platform can be capitalized under ASC 350-40 and amortized, while the subscription and usage are expensed. ASU 2025-06 takes effect for annual periods beginning after December 15, 2027. If you're paying a vendor or your own staff to set up integrations and workflows, ask your auditors whether that work qualifies.
A four-question test for your controller
Before you bill an AI run to a project, it should pass all four questions. The test is conservative on purpose, and you can use it today without buying any new software.
- Is it tied to one project number? If it can't be traced to a single job, it's indirect.
- Does the contract allow it? Look for ODC or technology-charge terms and any client AI clause. If the contract doesn't allow it, the cost stays in overhead or inside the fee.
- Can you produce the log? User, date, workflow, and cost by project number, in a form an auditor can test.
- Is it treated the same everywhere? Cost Accounting Standard 402-40 says, "All costs incurred for the same purpose, in like circumstances, are either direct costs only or indirect costs only." Decide by purpose, not by client.
Pass all four and it's a project cost. Miss one and it stays in the pool until you close the gap. Failing question three is a reason to fix logging, and you don't need to stop using AI on the job while you do.
Question four is where firms get into trouble. It's tempting to bill AI direct wherever a contract allows it and leave it in overhead everywhere else. Under CAS 402 and the FAR 31.202 and 31.203 definitions, that's double counting unless the circumstances actually differ, and government versus commercial work generally isn't accepted as a different circumstance. The fix is boring and it works: a one-page written policy, a log, and a credit-back to the pool for anything charged direct.
How AI hits your P&L depends on how the job is priced
When AI takes hours out of a deliverable, the saving shows up as margin on some contracts and as lost revenue on others.
| Contract | When AI saves hours | What to do |
|---|---|---|
| Lump sum or fixed fee | Saved hours become margin, and the AI cost is yours, inside the fee | Check whether new bids are still estimated on the old hours |
| Hourly or T&M | Saved hours become lost revenue, and AI goes on an ODC line if the contract allows it, otherwise in overhead | Move repeatable scopes to unit or task pricing where the client accepts it |
| Cost-plus fixed fee | The fee was set on estimated cost and profit is capped, so AI goes in the overhead rate, or ODC where allowable | Show AI in the rate proposal and ODC schedule so it gets negotiated |
On hourly and cost-plus work, getting more efficient can shrink revenue, and the industry went through the same thing when drafting tables gave way to CAD. The ACEC Research Institute and Virginia Tech study of lump-sum engineering services (September 2024) notes that some of this cost can go in the overhead rate, within regulatory and practical limits. From the vendor side, the firms we see keeping the savings are pricing repeatable scopes by the deliverable.
Public clients are already writing rules. Minnesota DOT's Standard IT-003 has consultants confirm their AI tools are approved before submitting non-public data. California's State Contracting Manual requires contractors to disclose generative AI used on deliverables that affect contract performance. Kansas policy P8200.00 requires disclosure of generative AI use and keeps state data out of AI tools without approval. ASCE Policy Statement 573 keeps responsibility with the engineer and says AI can't replace a PE's judgment.
Some firms do roadway work for clients who want no AI on the job at all, while their private MEP work is adopting it quickly. Add one field to the contract record, AI permitted, restricted, or disclosure required, so nobody bills an AI ODC to a client who never agreed to it.
AI moves your overhead rate from both sides
If AI sits in the pool and also reduces direct labor, the numerator goes up while the base goes down. Take an illustrative firm with $75.0M of indirect cost on $50.0M of direct labor, a 150.0% overhead rate. Add $1M of AI to the pool, let AI take 2% out of direct labor, and the firm has $76.0M on $49.0M, a rate of 155.1%. On an audited DOT rate, that's the number your clients see.
- Utilization falls if freed hours aren't refilled with billable work, so AI savings only reach the P&L when you have the backlog.
- Net labor multiplier rises on fixed-fee work, where the fee holds and labor falls. It stays flat or drops on hourly work.
- Overhead rate rises when AI is indirect, which is the strongest reason to move qualifying project AI into direct cost once it passes the four-question test.
Track four numbers per workflow, every quarter
Measure one workflow at a time, such as drawing QA/QC or submittal review, against the same kind of deliverable done without AI.
- AI cost per run, from the vendor's usage report, by project number. Ask every vendor for usage by user, workflow, and project.
- Hours before and after, from timesheets, on the same task codes, for the same kind of deliverable.
- Issues caught, from QA/QC logs and RFIs, so quality counts alongside hours.
- Expert review time. Engineers still check and stamp the work, so their time goes on the cost side.
Value per run = hours saved × loaded rate + rework avoided − AI cost − review time
Two cautions. Don't count hours nobody would have spent: if a full review wasn't going to happen anyway, count the result under issues caught. And don't leave out review time, because the engineer of record still owns the result.
The cheapest way to see where AI is used is an AI-assisted task code on your timesheets. It gives you before and after hours from the system you already run, and nobody has to fill out a survey.
Budget for 2027 in three layers
The firms furthest along budget AI in three layers that match the accounting buckets, so the budget and the books tell the same story. These examples come from our customer conversations this fall, with names removed.
- Fixed: the base platform. Seats and any included usage, known at signing and owned by IT or operations. At one global design firm, everyone gets $25 a month for general use, booked to a software cost center.
- Variable and direct: the project pool. AI estimated in each project budget by workflow and phase, like printing or CADD, and owned by the project manager. At that same firm, the project manager approves a pool when the project is set up and it posts to the project code, and finance splits the export so software cost and project cost land in different places. An architecture firm's first pricing question to us was whether AI could be a reimbursable.
- Variable and indirect: the firm pool. Training, proposals, pilots, and unbilled client work, owned by a named leader with a cap and an end date. Several mid-size firms fund year one from the innovation budget. Once usage grows, plan the move to project budgets so project cost doesn't hide in overhead.
A placeholder beats a zero. One design firm leader told us the worst outcome is a zero in the budget and a request to sign a contract three weeks after the budget closes. Put a number in the 2027 budget now, then run one quarter on demand before you sign an annual usage commit. Vendors, including us, offer better terms for an annual pooled commit, and that's reasonable once you know your run rate. Before then you're guessing.
What didn't work
These six mistakes cost firms time or money in pilots and rollouts we ran this year. Each one was cheaper to avoid than to fix.
- Trusting the drawing as complete. A reviewer gave us a sheet chosen because it had known problems, and the first code-compliance pass came back clean. The agent checked what was on the drawing and missed rooms with no occupancy tags. We fixed the workflow, and the takeaway for any firm is to score pilots on a set with known issues and to check completeness before code.
- Mixing setup spend with run spend. A contractor's usage from building its first workflows came out of the production pool and threw off its cost per run. Put implementation and run budgets on separate contract lines, which also lines up with the capitalization bucket.
- Month-end request floods. People on small monthly allowances send top-up requests at month end, and the approver can't tell which pool or project each request is for. Route each request to the pool owner, with the project attached.
- The Ferrari problem. Given a choice, people pick the most capable model for every task, including the ones a cheaper model handles well. Set a default model per workflow, with a named owner.
- One person holds the keys. One office routed every run through a single person. Spend stayed predictable, and the value was capped at what that person could run. Give teams project budgets and caps, then open access.
- Buying before picking the workflow. A civil firm's engineers liked the demo but couldn't name the workflow they'd measure, so nobody could show a return. Record current hours on one workflow before you sign.
Eight controls to require from any AI vendor
These apply to every AI product at your firm, including the general chat tools your staff already expense. Ask every vendor for this list in writing, and that includes us.
- Spend by project number, reported by project, user, and workflow, using your ERP's project numbers.
- Hard limits at the firm and user level that pause usage when they're hit. An email alert at 11 PM on a Friday doesn't stop a background review of three drawing sets.
- A cost estimate before a run, for large jobs, before anyone commits to them.
- An approval path to request more budget and route the request to an approver.
- Export and API access, so you can pull usage data into your ERP every month without retyping it.
- Visible model costs: what the underlying model costs, and what the vendor adds on top.
- Invoice and PO billing, with invoices that reconcile line by line to the usage report.
- Data isolation, so your files stay in your firm's environment and aren't used to train models.
Then close AI spend monthly, like any other project expense. Require the project number when the work starts, export spend by project, import it into Vantagepoint, Costpoint, or whatever you run, and credit the pool for anything charged direct. You probably already do this for reprographics or a travel card feed. Whatever can't be assigned stays in overhead, and that unassigned number is your to-do list each month.
Here are our answers. In Nomic, each project carries your ERP project number, and admin analytics shows spend by project, member, and workflow, with a filtered CSV export and an analytics API that returns the project number on each usage event. Firm and user spend limits pause usage when reached, large workflows show a cost estimate first, and requests for more budget land in the approver's inbox. Usage is billed at the underlying model rates with no markup, invoice and PO billing is available on Enterprise, and each firm runs its own environment, with model providers working under zero-data-retention terms. The details are on our pricing and security pages and in the changelog.
Five things to do before the 2027 budget is final
None of these need new software.
- Sort last quarter's AI invoices by project, user, and tool. Whatever you can't assign is your starting overhead number.
- Write a one-page AI cost policy that says which uses are direct, which are indirect, and what evidence you keep. Walk your auditor through it before year end.
- Add an AI-assisted task code so timesheets show where AI is used on each project.
- Back-test one workflow on a finished job where you know the answer, then track the four numbers on live work for a quarter.
- Flag AI terms on your largest contracts as permitted, restricted, or disclosure required, and record that on the contract.
The policy is what protects you in an audit, and the invoice sort tells you how big the problem is today. If you only have time for one before budget season, sort the invoices.
If you want to see where the cost data comes from on your own work, book a 30-minute demo and bring your controller. We'll run a review on your drawings or a sample set and show the cost estimate before the run, the cited findings, and the project-level usage export your ERP would import.
Sources
- Deltek, 47th Annual Deltek Clarity Architecture & Engineering Industry Study, released May 12, 2026 (896 firms).
- KPMG, Global AI Pulse Q2 2026, June 2026 (2,145 senior leaders at organizations with $50M+ revenue, 20 countries).
- AASHTO Uniform Audit & Accounting Guide, Appendix B, Internal Control Questionnaire for Consulting Engineers, item H.8.
- Cost Accounting Standard 402-40, and the FAR 31.202 and 31.203 definitions of direct and indirect cost.
- West Virginia Division of Highways, Consultant Audit Guide, on direct computer charges and pass-through billing rates without profit.
- FASB ASC 350-40 and ASU 2018-15 on cloud computing implementation costs, and ASU 2025-06, issued September 2025.
- ACEC Research Institute and Virginia Tech, Provision of Engineering Services on a Lump Sum Basis, September 2024.
- Minnesota DOT Standard IT-003; California DGS State Contracting Manual GenAI disclosure provisions; State of Kansas Policy P8200.00, section 9.2.7; ASCE Policy Statement 573, adopted July 18, 2024.
- Nomic unit costs from docs.nomic.ai Models & Pricing. Field results come from Nomic customer projects and conversations in 2026, with firm and people names removed.
This post is not tax or audit advice. Talk to your auditors and tax advisors before changing how you account for AI.
