Your AI teammate has a spiky profile. So do your best people

We spent twenty years learning to build teams around people whose strengths and gaps cluster unevenly. That turns out to be the exact skill AI collaboration requires.

When a client tells me "AI can already do 90% of this," arguing about output volume puts your retainer on weak ground. The stronger move: map where each contributor, human or machine, is exceptional and where each one falls off. Use the four-part envelope map below to assign the work and name the final 10% of judgment that stays yours.

In 1999, Thorkil Sonne's three-year-old son Lars was diagnosed with autism. Sonne was a technical director at the Danish telecom TDC, and he did what a technical director does. He went looking for the data. The research gave him a deficit model. At his kitchen table, he saw a boy who could reproduce a European road atlas from memory.

Five years later he mortgaged the house and started Specialisterne, a software testing firm staffed almost entirely by autistic people.

The pitch to clients was not charity. It was that a tester who can hold a 400-case regression suite in working memory and stay on it for six hours without drifting is simply better at testing than a neurotypical generalist who gets bored at case ninety.

I have thought about Sonne a lot this year, and not because of hiring.

The machine that holds a tolerance it can't question

I have been running AI across every part of my work: drafting, research, financial modeling, board prep, translation.

What keeps surprising me is not that it is good or bad. It is how unevenly those two things sit next to each other.

Last month it reconciled a compensation model across three entities and caught a rounding inconsistency I'd carried for two quarters. An hour later, in the same session, it produced a paragraph that contradicted something I'd told it four messages earlier. Not a hard thing. An obvious thing, the kind a competent second-week hire would never miss.

A five-axis CNC mill will hold a tolerance of two thousandths of an inch all day, on the ten-thousandth part, long after a machinist's hands would have started to wander.

It is not better than a machinist in some general sense. It is extraordinary inside its envelope and blind outside it. It will cut a part to spec with beautiful precision, and it will never once look up and tell you the part is designed wrong.

That is the shape. Spikier.

Comparing people to machines has an ugly history, so let me name it

Here is where this essay gets awkward, so let me put it on the table rather than write around it.

Comparing people to machines is one of the oldest ways to dehumanize them, and comparing autistic people to machines has a particularly ugly history.

If this piece reads as "AI is like autistic people," I have written it badly and you should stop here.

The claim runs the other way. The people I'm talking about are the teachers. AI is the subject.

A generation of managers learned something real about collaborating across very different cognitive profiles, mostly without calling it that. Those hard-won practices are the most directly transferable management skill we have for the thing now sitting in our workflows.

The transfer runs from human to machine. Not back.

One more piece of housekeeping. Asperger's stopped being a separate diagnosis when the DSM-5 folded it into autism spectrum disorder in 2013, and the name carries baggage from Hans Asperger's conduct in Nazi-era Vienna.

I use autistic rather than "person with autism" because that is the preference most autistic adults I have worked with have stated. If I have that wrong for you, tell me and I will use your word.

The envelope: four practices that transfer

Everything I have learned about collaborating across spiky profiles reduces to four practices. They came from managing humans. They apply, almost without modification, to managing AI.

1. Draw the envelope before you delegate. Spiky performance is only unpredictable if you have not mapped it. Once you know the strengths and gaps, the variance stops being mysterious and becomes a routing problem.

   A collaborator may be exceptional at sustained structured analysis and weak at reading unstated social context. Most of the frustration I hear about AI is really the cost of skipping that distinction before the work gets assigned.

   One approach that worked on my own engagements: list each recurring task, then mark where you, each colleague and each AI tool are exceptional, and where each one falls off. Use observed work, including the obvious misses.

2. Supply whatever sits outside the envelope. The most reliable fix for a context miss is to stop treating context as ambient.

   Say the thing you assume everyone knows. Write down the constraint that currently lives only in your head.

   With human colleagues this felt at first like over-explaining, and then it felt like basic professionalism. The AI version is the same lesson with a shorter feedback loop and no social cost for repeating yourself.

   Add the required context beside each task in your map. That turns "AI needs better prompts" into a specific operating question: which missing constraint caused this output to fail?

3. Never downgrade the whole machine for one miss. This is the expensive error, and I watched managers make it with people for two decades.

   Someone misses a social cue in a client meeting and their analytical work quietly stops getting taken seriously. One gap gets read as a general ceiling.

   We do exactly this to AI. It hallucinates one citation and the team concludes it's useless for research, which is a bit like scrapping the mill because it can't read a drawing.

   Record the miss against the task and the condition it happened in. That gives you a specific control to add without discarding work the contributor still does well.

4. Staff for complementary envelopes, not for an average. The team that wins isn't the one where everyone is adequate at everything. It's the one where the coverage map has no holes.

   That principle predates AI by a century, and it is the single most useful thing to carry into a workforce where some of the envelopes now belong to software.

   Route each task to the contributor with the strongest observed fit, then list what is left uncovered: judgment, missing context, review. That uncovered area is where the engagement still needs you.

What the accommodation literature already figured out

There is a body of practice most technology leaders have never read, and it is sitting right where we need it.

SAP launched its Autism at Work program in 2013. Microsoft started its autism hiring program in 2015, during the years I was there.

I remember the internal reaction being a mix of genuine enthusiasm and quiet skepticism about whether the interview loop could be changed at all.

Robert Austin and Gary Pisano wrote the case for it in Harvard Business Review in 2017. Their argument was not that neurodiverse hiring is kind. It was that standard management practice is tuned to a narrow profile and quietly discards the value outside it.

The accommodations those programs adopted now read like a prompt engineering guide:

• Give instructions in writing rather than implying them in a hallway.

• Make evaluation criteria explicit instead of assuming they are understood.

• Replace the unstructured interview with a work sample. The interview measures social fluency and calls it competence.

• Structure the environment rather than asking the person to absorb the ambiguity.

Every one of those produces a better AI workflow, as written, without translation.

Temple Grandin has spent forty years arguing that different minds solve different problems, and that a world designed for one cognitive style loses the problems only other styles can solve.

She was talking about people. The sentence survives the substitution better than it should.

Orchestration, not prompting, is the skill that holds its value

Here is what spiky profiles and complementary envelopes add up to for anyone building a company or defending a client retainer right now.

Prompting and model selection buy you months. A model decision has a shelf life you can measure on a calendar.

The scarce skill is orchestration: knowing where each contributor is exceptional, where each one falls off, how to route the work, and how to make context explicit at every handoff.

That is a management competency, not a technical one.

The people who already excel at it have usually spent years working alongside a colleague whose mind worked differently from theirs, learning together to produce something neither could produce alone.

We used to call that inclusion and file it under HR. We were building the core competency for AI-era leadership and filing it in the wrong drawer.

For a fractional CMO, that final 10% is judgment: what matters, who handles it, when an output is wrong, what reaches the client.

You can make that value visible in one sentence:

AI handles [named outputs] inside a defined envelope. I earn the retainer by deciding what matters, supplying client context, catching failures and taking responsibility for what ships.

Name the tradeoff before you claim the upside. Once you claim the judgment layer, you own the routing decisions and the quality of the final work, including the parts a model produced. That accountability is what the retainer buys.

Go find the manager in your organization who is quietly brilliant at this. Watch what they do at each handoff. Then teach those mechanics to everyone else before the market prices them properly.

If you want the underlying argument in long form, it is the spine of Syntropy: as production gets cheap, the premium moves to judgment about what matters and who should do what.

Start with the chapter on editorial judgment, then run your own team through the envelope exercise above and mark where the coverage map still has holes.