Every process in your company. And the few that shouldn't need you.
You've seen the posts promising an AI that runs a company with nobody in it. This is the version that survives contact with a real business: a map of every operational process, honestly labelled — running itself, running with a person, or untouched. In most companies the third category is the biggest, and that's the correct result.
Nobody is running a real company with no humans in it. Those demos are impressive and they're demos. What is true, and worth money, is that roughly one process in five can run unattended — and knowing precisely which five is the entire job.
Select any node on the map to see what
it is, what state it's in, and what
would have to be true to change that.
Or
filter by department above.
Why most of that map is grey, and should be
A map where everything glows green is a rendering, not a business. Four reasons a process stays human.
Judgement with consequence
Approving a refund above a threshold, rejecting an application, deciding what to tell an unhappy customer. A machine can assemble the case; someone has to own the decision, and in regulated work someone must be nameable.
Volume too low to matter
Roughly half the processes in any company run a handful of times a month. Automating them is technically possible and economically pointless — the build costs more than the annoyance.
The inputs aren't written down
Where the knowledge lives in someone's head, there's nothing to build on. That's a documentation problem first, and no amount of orchestration solves it.
It changes too often
A process that's rewritten every quarter shouldn't be encoded. Automating it buys rigidity at the exact moment you need flexibility.
The value of the map isn't the green nodes. It's knowing which grey ones to ignore. Most companies burn their first AI budget on a process that was never going to pay back, and the reason is that nobody ranked them before starting.
Where the wins usually are, and roughly how big
Patterns across mapped companies. Yours will differ — that's the point of mapping rather than assuming.
High volume, answers that already exist in writing, and a clear quality bar. Suggested replies and triage are the most reliable first win in almost every company I've looked at.
2–4 processesstrong case
Invoice extraction, matching, exception routing. Unglamorous, very repetitive, and the volume means small per-item savings compound quickly. Approval stays human.
3–5 processesstrong case
Exception handling and status chasing usually automate well. Anything depending on a third party's data quality is harder than it looks and often stalls there.
2–4 processesdepends on data
Research, note capture and CRM hygiene are genuine wins. Anything touching the actual conversation is usually worse than the human and quietly damages relationships.
1–3 processesnarrow wins
Policy answers work well. Almost everything else is low volume or too sensitive to route through a machine, and the map usually recommends leaving it alone.
1–2 processesweak case
Content generation is where most companies start and where returns are hardest to measure. It's also the one place your team is probably already using tools without you.
0–2 processesusually skip
Three weeks, and you keep it whatever happens next
Interviews
Two hours with each department lead and, more importantly, with the people actually doing the work. Their description differs from the manager's every single time.
Volume and cost
Each process gets a frequency, a loaded cost and a data-readiness rating. This is what turns a diagram into a ranking.
Ranking and handover
The map, plus a ranked shortlist with estimated cost and payback for the top five — and an explicit list of what not to bother with.
The map is a deliverable, not a sales meeting
Plenty of companies take the map and act on it themselves. That's a fine outcome and it's priced accordingly.
Every process, ranked, with the ones to ignore marked.
$9,500Three weeks · credited against a build
- Interactive map of your operations
- Volume and cost against each process
- Top five ranked with payback estimates
- An explicit "don't bother" list
- Yours to act on without me
The top-ranked process, built and running.
$18k – $40k6–8 weeks · map credited
- Accuracy bar you set, measured
- Runs in your infrastructure
- Parallel run before cutover
- Monitoring and runbook
- Map updated with the real result
Work down the ranked list on a monthly rhythm.
from $6,500/moCancel any month · 30 days notice
- One process shipped per month
- Everything already live kept running
- Map maintained as the live view
- Your engineers can pair throughout
- Stops when the list stops being worth it
What the map is allowed to conclude
That you shouldn't build anything. It happens — usually when volumes are lower than assumed, or when the processes that hurt most are the judgement-heavy ones that should stay human. You'd have a ranked map, a costed picture of your own operation, and a written recommendation to stop.
The reason that's an acceptable outcome for me is that I'd rather do three good builds a year than six that disappoint. A map that says no is cheaper for you than a build that underdelivers, and better for me than a reference I can't use.
Model and infrastructure costs sit in your own accounts throughout. Not marked up.
What people ask about the map
I've seen demos of AI agents running an entire company. Why can't we have that?
Because those demos aren't running a company — they're running a scripted scenario where every input is well-formed and nothing has consequences. It's a genuinely impressive engineering artefact and it isn't a deployment.
The gap isn't capability, it's accountability and edge cases. Real operations are mostly exceptions, and exceptions are exactly where unattended systems fail expensively. What you can have is the fifth of your processes that are high-volume, well-documented and low-consequence running with little or no supervision. That's a real and valuable thing; it just doesn't make a good video.
Can't we just map this ourselves?
You can, and some companies should. What you'd be paying me for is the ranking rather than the drawing — knowing which processes are technically feasible today, roughly what each would cost, and which ones look attractive but reliably disappoint.
If you have someone internally who's delivered a few of these, they'll do it better than me because they know your business. If you don't, the ranking is the part that's hard to get right from the inside.
Will this be used to justify cutting jobs?
That's your decision and I'd want it stated at the start, because it changes what gets built and how success is measured. In most engagements headcount holds and throughput rises — the tedious gathering goes, the judgement work stays.
What I won't do is produce a map framed as a redundancy plan while telling your staff it's about efficiency. The people being interviewed will work out which it is faster than leadership expects, and the interviews stop being truthful the moment they do.
How accurate are the estimates on the map?
The volume and cost figures come from your own systems and timesheets where they exist, and from observation where they don't — those are reasonably solid. The feasibility ratings and payback estimates are judgement calls, and I'd treat them as ±40% until a pilot measures one.
That's why the map ranks rather than promises. Its job is to stop you starting with the wrong process, which is where most of the money gets lost.
Our processes aren't documented. Is the map still possible?
Yes, and the map is usually more valuable in that situation — it's often the first time anyone has seen the whole operation written down in one place. Several clients have got more from that than from anything built afterwards.
It does change the findings though: undocumented processes rate lower on feasibility, because there's nothing for a system to learn from. Expect the map to recommend writing things down before automating them.
How long does the map stay accurate?
Six to twelve months for a stable company, less if you're growing quickly or changing systems. The ranking ages faster than the map itself, because what's feasible shifts as models and tooling change.
The ongoing option keeps it current as a live view rather than a document, but that's only worth paying for if you're actually working down the list.
Eight years of systems that ran unattended, which is why I'm cautious about promising it
I'm Bahman Shadmehr. Automated trading platforms placing orders on the Texas wholesale power market and US equities, a national payment gateway I helped decompose while live, and the pipelines and monitoring underneath all of it.
I've actually built the thing those posts are gesturing at — systems that act with no human watching. That's precisely why I'm careful about it. The engineering isn't the hard part; knowing which decisions are safe to hand over, and building the alerting for when you were wrong, is where the years go.
You deal with me from the first interview to handover.
Python, Go, Kubernetes, Docker, AWS, GCP, Postgres, Redis, Prometheus. Remote, UTC+3. 76 technical posts · LinkedIn
Name the three processes your team complains about most
I'll tell you within a day which of them are likely to be worth automating and which are complaints about something else. That's a useful answer either way, and it costs you an email.
info@bshadmehr.meUseful in a first message
- Headcount, and which departments exist.
- The three most-complained-about processes.
- Whether anything is automated today.
- What you've already tried, including failures.
- Whether the goal is capacity or cost.