The Do-Not-Automate List

Recently I saw a presentation on AI that had a slide on things you shouldn't automate, with the focus being on parts of your job that you love. Often these are the bits that require novel problem solving or using your own judgement based on your experience. As someone who works in AI, I'm equally passionate about things that should remain human, so this got me thinking about all the reasons you might not want to automate something.

Every business leader is thinking about what they can automate with AI, and most teams approach it the same way. They gather the leadership team, open a use case listicle, and start brainstorming everything AI could possibly do. Six months later they have a long document, a stalled pilot, and a team quietly worried about what the exercise means for them.

There is another way in, and it starts from the opposite end. Instead of asking what AI could do, ask what should stay human.

The Do-Not-Automate List is not a replacement for use case hunting. It is a complementary lens, and a surprisingly productive one. Drawing the human line first makes everything on the other side of the line easier to see, and safer to automate. The counterintuitive payoff is that deciding what to keep human ends up surfacing more genuine automation candidates.

Why the stakes are higher for services firms

If you run an agency, a consultancy, a recruitment firm or a professional services business, your product is judgment and trust. Automate the wrong thing and, in looking to add efficiency, you end up automating what differentiates your business.

Your clients run the same AI tools you do. They know what a generated deliverable looks like, and if your work reads as AI-average, the taste and the costly signals your fee depends on are exactly where the value dies. Researchers at MIT Sloan draw a useful distinction here between tasks and trust: AI output works well as raw material, but clients are paying for the finished judgment sitting on top of it. In finance, where the cost of getting it wrong is existential, Do-Not-Automate lists are already standard practice.

There is a second reason to run the exercise, and it has nothing to do with clients. It gives your people agency. When each person on the team identifies the parts of their job they want to preserve and the parts they would gladly hand off, AI stops being a threat imposed on them and becomes a tool they have shaped. Most adoption resistance dissolves before it has a chance to form.

The six tests

Run every workflow in your business through the six tests below. Anything passing a test stays human. Anything failing all six becomes a candidate for automation. We'll come back to what to do with those candidates at the end.

Test 1: Accountability is the product

Some work exists because a client needs an accountable adult who owns the outcome. Signing off audit opinions or legal advice, presenting to the client's board, approving scope and pricing exceptions, and delivering bad news all belong to humans, permanently. No agent can carry accountability on a client's behalf, and no client will accept one trying.

What the test clears for AI: the workpapers underneath the opinion, the assembly of the board deck, and the pricing scenario modelling behind the exception. Each is a focused job with a clear input and a clear output, which makes each one ideal agent work.

Test 2: The relationship is the value

Retention is the business model of every services firm, and clients churn when they feel processed rather than known. Quarterly business reviews with your top accounts, partner calls with no agenda, renewal conversations and the work of saving a wobbling client stay human forever.

What the test clears for AI: everything wrapped around those conversations. Meeting preparation and briefing documents, CRM hygiene, post-call summaries and follow-up drafts are exactly the hours partners resent, and none of them require the partner.

Test 3: Taste is your differentiator

AI regresses to the mean by design, and a firm built on AI-average opinions becomes a commodity firm. The creative concept, the consultancy's contrarian point of view and the headhunter's judgment call on the final shortlist are the reasons clients chose you over the cheaper alternative, so they stay human.

What the test clears for AI: the research underneath the point of view, competitive teardowns, first-pass longlists and asset production once the concept is set. Taste directs the work, and agents produce the volume beneath it.

Test 4: Errors are irreversible or unbounded

Some mistakes cannot be walked back. Filing deadlines and regulatory submissions, the final go or no-go on a deliverable, and anything sent under the client's name to their customers need a human in control, because one bad automated send can end a seven-figure retainer.

What the test clears for AI: the drafting of all of it, provided a human gate sits in place before anything leaves the building. The test is about placing checkpoints deliberately rather than avoiding automation altogether. One caution worth taking seriously: human-in-the-loop fails when humans become rubber stamps, so put your checkpoints where they carry real weight, not on every step.

Test 5: Doing it is how you learn

Services firms run on apprenticeship. Juniors sit in on discovery calls, write the first drafts of strategy documents and do enough of the analysis to build judgment, because the grind is the training. A firm which automates all of its junior work will find itself without a strong talent pipeline for senior roles.

What the test clears for AI: judgment-free volume work such as data entry, formatting and transcription, which never taught anyone anything. Protect the formative work and hand over the rest.

Test 6: The signal is costly

Sometimes the effort is the message. The handwritten note after a big win, the partner flying out for the pitch and the genuinely bespoke proposal all work because the client can see what they cost you. Clients can smell AI personalisation from the first line, and a costly signal produced cheaply is worse than no signal at all.

What the test clears for AI: signals nobody reads as effort. Scheduling, reminders and standard onboarding documents carry no message, so hand them over without hesitation.

The litmus test

If the six tests feel like a lot to hold in your head, one line covers most cases.

If you can name the trigger, the rules and the output, it is agent work. If you cannot, it is human work.

The line maps directly onto how well-built agentic systems operate: a trigger fires, an agent runs, an action follows. A deal closes and a handover document is produced. A document lands and it is classified and routed. A ticket escalates and a briefing is prepared. Each is a discrete workflow expressible as: given X, produce Y. The moment you cannot define the trigger or the rules, you are describing judgment, and judgment is what your clients are paying for.

The payoff: the list is short, and the backlog builds itself

Here is what surprises most teams who run the exercise. The Do-Not-Automate list ends up short. Once the human line is drawn, everything on the other side becomes a candidate for a focused specialist agent, each one doing a single clearly defined job with only the context it needs. The best AI is the smallest AI, and the six tests hand you a ready-made map of where the smallest useful agents should go first.

The commercial logic is straightforward. AI absorbs the hours clients resent paying for, such as research, formatting and administration, while your people keep the hours clients actually want to pay for: judgment, presence and accountability. Margins improve on the work clients never valued, and the work they do value gets more of your team's attention.

One final piece of advice: run the six tests with your team, not on them. Each person maps the work they want to keep, which is the work making the job theirs, against the work they would gladly hand to an agent. The automation backlog builds itself, pre-approved by the people it affects.

Then start small. Pick two or three agents delivering standalone value and stack from there. As one Rightbrain customer at Rocket SaaS put it: "We started with one agent where we knew we'd see value. Now we're stacking different use cases."

Where to start

Or run it with us. In a free AI audit, we map your workflows together, draw the human line, and identify the two or three high-impact agents worth building first. There is no pitch and no demo request, just a working session on where AI genuinely fits in your business. If there is a fit, we will build a demo together.