AI Agents for Professional Services: Delegate by Client Risk, Not by Hours
In October 2025 Deloitte Australia agreed to partially refund the government for a report on a contract of about A$440,000, after reviewers found footnotes and references that did not hold up. The revised version disclosed that part of the work used a GPT-4o model. It is the clearest recent picture of what goes wrong with AI agents for professional services.
The model was not too slow or too expensive. It sat in the one workflow where a wrong sentence is read by a paying client, and checking it was nobody’s named job.
Delegate by the cost of a wrong output
When a firm picks its first agent, it usually picks by hours. Delivery holds the most billable time, so delivery gets the agent. I think that order is backwards.
The adoption numbers explain why. In the Thomson Reuters 2026 AI in Professional Services report, 15% of professionals say their organization already uses agentic AI tools and another 53% are planning or considering it. Only 18% say their organization tracks ROI. A lot of firms are buying agents with no way to see whether they pay back.
That is the setup Gartner describes when it predicts that over 40% of agentic AI projects will be canceled by the end of 2027, citing rising costs, unclear business value and weak risk controls. Hours saved is the metric that gets a project cut, which is the argument in AI agent ROI: stop measuring demos, start measuring outcomes. A better first question is what happens when the agent is wrong.
A professional services firm sells judgment and bills for it, so the agent question is really a question about where judgment gets applied and where it is merely assumed. A SaaS company can ship a wrong answer, patch it and move on. A firm that ships a wrong answer has spent the one asset clients are paying for, and a refund is the cheap version of that.
Score two things, then pick a mode
For each workflow I would ask two questions.
Judgment risk: how much professional judgment does it take to notice the output is wrong? A reminder to submit a timesheet needs none. A reading of a tax position or a contract clause needs a lot.
Client impact: if a wrong output reaches the client, what breaks? A late meeting note is an annoyance, and a wrong number in a deliverable is a refund.
Low on both means the agent acts and a person samples the results. High judgment risk with modest client impact means the agent assists: it drafts, a professional approves. High client impact, whatever the judgment score, means the agent escalates. It gathers the evidence, flags its own assumptions and stops, and a named professional decides.
Anthropic’s guide to building effective agents makes a similar point about design: an agent can pause for human feedback at checkpoints or when it hits a blocker. The scoring above only decides where those checkpoints go in a service firm, and it forces the conversation about who owns each one before the agent goes live rather than after the first complaint.
Where the four workflows land
| Workflow | Default mode | What the agent does |
|---|---|---|
| Administration | Act | Scheduling, time-entry chasing, filing, internal meeting notes |
| Growth | Act, then assist | Prospect research and list building; drafts for a partner to approve |
| Client operations | Assist | Intake, document collection, status updates drafted for review |
| Delivery | Escalate | Source gathering, first-draft sections, assumption flags; a professional signs |
The table is a default, not a rule, and a firm can move a row once its own data supports it.
Administration goes first because the cost of a mistake is internal and cheap to catch. It is also the cleanest place to practice logging what the agent did, which what to automate first argues should come before anything fancy. A firm that cannot say which timesheet reminders the agent sent last week has no business letting it draft a client memo.
Growth comes second, and I would still let people approve anything sent under a partner’s name. Research and list building are low risk. A badly personalized first email is not, because it spends a relationship you have not built yet. The wider gap between pitch and reality in sales agents is covered in best AI agents for business.
Client operations is assist-only for me. Document collection and status updates feel routine. They are also the messages clients read most often, and a status update that says a deliverable is on track when it is not creates a problem that no later draft can fix, so a human reads them first. Once the rejection rate on these drafts stays near zero for a quarter, I would move the lowest-stakes categories, such as meeting confirmations, to act mode and leave the rest.
Delivery is where I would be strictest, and where I expect the most pushback. If an agent drafts any part of what the client pays for, I would want a named professional to check every cited source against the original, and a record of how often the reviewer rejects the draft. The Deloitte episode is what happens without that record.
The objection that delivery is where the margin is
It is true, and it is the best argument against my ordering. Delivery is the biggest cost line, and 77% of professionals in the Thomson Reuters survey expect agentic AI to be central to their workflow by 2030. Firms that wait will be undercut.
My answer is that the margin does not move because someone drafts faster. It moves when the delivery model changes, which is the point of productizing AI services, and a delivery model you cannot audit is not a product. Starting with administration and growth builds the logging and the review habit that delivery then depends on. Skipping them does not make the firm faster, only less able to explain a mistake.
Measure the reviewer, not only the agent
The agent’s output is the easy part to count.
Most reporting on agents counts what the agent produced. The number that tells you whether a mode is right is the reviewer’s rejection rate: how often the professional changes or discards the draft, and what kind of error they caught.
If a growth draft is rejected one time in twenty, assist mode is working and a move toward act mode is reasonable. If a delivery draft is rejected one time in three, the agent is creating work, and the hours you thought you saved went to the person fixing it. That second case is invisible in a tool-adoption dashboard, which is why a survey where only 18% track ROI should worry anyone about to scale.
What I would do with this next quarter
I would pick one administration workflow and one growth workflow, set the mode for each in writing, and measure rejection rates for 60 days. Delivery stays in escalate mode until a named reviewer exists for every output, because a firm that cannot name the person who signs has already answered the question of who is accountable when the client calls. That is a governance gap, not a tooling gap. The mode table should be short enough to print and argue about in a partner meeting.
If you want to start with the growth side, I run a free growth audit at app.momentumnexus.com that shows which of your pipeline steps an agent can take over first.
Ready to Scale Your Startup?
Let's talk about the agents we could run for you to put these strategies to work.
Schedule a Call