In short
- Adding AI to each step preserves the shape a process got from constraints that no longer hold.
- Three of those constraints have collapsed: the cost of attention, the cost of reading and writing, and the cost of checking quality.
- Five redesign moves: delete, collapse, invert, re-gate, instrument.
- Document the change as a before/after ledger, with a baseline, or you will never know whether it worked.
- Then answer the question most teams avoid: what does the freed capacity become?
The warning is 35 years old
In the July–August 1990 issue of Harvard Business Review, Michael Hammer published Reengineering Work: Don't Automate, Obliterate. His argument was that companies facing slow cycle times, high error rates and sluggish customer response would not fix them with incremental improvement — and that laying new technology over an existing process mostly preserves the problem while adding expense.
Everything that made that warning easy to ignore in 1990 is stronger today. Building is faster, demos are more convincing, and the vocabulary is more exciting. So teams add a model to step four of a nine-step process, save eleven minutes, and report a transformation.
There is also a measurement reason to expect a gap between capability and results. In the NBER working paper Artificial Intelligence and the Modern Productivity Paradox, Erik Brynjolfsson, Daniel Rock and Chad Syverson examined why impressive AI capability had not shown up in productivity statistics. Of the four explanations they considered, they concluded that implementation lags were likely the biggest contributor: like other general purpose technologies, AI's full effects arrive only as waves of complementary innovations are developed and implemented, and the required organizational changes and new skills behave like an investment in intangible capital.
Translated for an operator: the model is the cheap part. The redesign is the investment.
The three constraints that shaped your current process
Every workflow you inherited was optimized against conditions that were true when it was designed. Name them, and the redesign becomes obvious.
1. Human attention was scarce and expensive. So work was batched ("we process applications on Thursdays"), queued, and rationed by priority. Queues exist because attention was the bottleneck. What holds now: the reading, sorting, drafting and lookup portion of attention is available continuously and priced per unit.
2. Reading, looking things up and writing took minutes each. So steps were bundled: one person "owns" the case end to end, because handing it over cost more than doing it. And research was skipped for anything but the biggest accounts. What holds now: a lookup or a first draft takes seconds, so the economics of bundling change.
3. Checking quality was expensive. So it was done by sampling, at the end, by a supervisor. What holds now: a check can run on every record, inside the process, before the damage.
If you keep the batches, the bundles and the end-of-line inspection while adding AI to the steps, you have a faster version of a process built for a world that no longer exists.
The five redesign moves
Take one workflow you have already mapped — and map it first, using the method in process mapping before automation — then apply these five in order.
1. Delete
Find the steps that exist only to compensate for slowness, bad data or missing visibility. Status-chasing messages. The weekly meeting where people read numbers aloud. The re-keying of data between two systems. The reminder to send the reminder.
These are not candidates for automation. They are candidates for deletion, and they are usually the largest share of the wait time on your map.
2. Collapse
Find steps that were separate only because different people did them, or because one person could not hold it all at once. Read, classify, summarize and route can frequently become one step with one structured output.
Rule: collapse only where you can still validate the result. Two steps with a check between them beat one step you cannot verify.
3. Invert
Most processes are reactive: something arrives, then work begins. Invert it. Do the work before it is requested, because the marginal cost of preparing is now low.
Instead of researching an account when a rep picks it up, arrive with the research already attached. Instead of writing the report when someone asks, have it written and waiting with the numbers already pulled. Inversion is where AI-first workflows feel qualitatively different rather than just quicker.
4. Re-gate
Move the human from doing the work to deciding about the work. The gate should sit where a wrong answer costs money, trust or compliance — and nowhere else.
Re-gating is what keeps an AI-first workflow trustworthy and is the difference between a Doer and a Director in MitHub's capability ladder. The person still owns the outcome. They stop owning the typing.
5. Instrument
Every step leaves a signal: what came in, what was decided, what went out, how long it took. Without instrumentation you cannot tell an improvement from a coincidence, and you cannot debug a silent failure.
This is also the only way the fourth move stays honest. A human gate with no record of what was approved is theatre.
The before/after ledger
Document the redesign as a ledger, not a slide. One row per step of the old process.
| Old step | Why it existed | Does that still hold? | New design | Who decides |
|---|---|---|---|---|
| Rep researches the account | Data was expensive to gather | No | Research attached before assignment | Rep judges fit |
| Ops re-keys form data into CRM | Systems were not connected | No | Direct write, validated fields | Nobody — logged |
| Weekly status meeting | No shared visibility | No | Live view + exception alert | Manager on exceptions |
| Manager approves discounts | Wrong answer costs money | Yes | Unchanged, with context prepared | Manager |
That fourth row matters as much as the first three. An AI-first redesign is not a purge of human judgment; it is the removal of everything that was never judgment in the first place.
The discipline of drawing the current state before the future state comes from lean practice — the Lean Enterprise Institute describes value-stream mapping as exactly that pair of maps, current and future, aimed at seeing the sources of waste in the whole flow rather than one step.
A worked example (hypothetical)
Imagine a 12-branch home services company handling quote requests.
Before: a form arrives in a shared inbox. A coordinator checks it during their inbox block, twice a day. They look up the address, guess at job size, assign a branch from a printed territory list, and email the branch manager. The manager replies with availability. The coordinator emails the customer. Average elapsed time: about two days. Touch time: about 20 minutes.
After the five moves:
- Delete: the manager-reply email round trip, which existed because availability lived in someone's head rather than a calendar.
- Collapse: validation, address lookup, job-size classification and territory assignment become one step with a structured output (branch, job tier, confidence).
- Invert: the quote draft and the branch's next three open slots are prepared the moment the form arrives, not after a human reads it.
- Re-gate: the branch manager approves or edits the quote. Low-confidence classifications route to the coordinator instead of guessing.
- Instrument: time-to-first-response, share of low-confidence routes, and quote-to-job conversion by branch.
The coordinator role does not disappear. It changes from routing to watching the exceptions and improving the classifier — which is a more valuable job and a harder one to replace.
Redesigns of this shape can run at real scale: MitHub's pioneers have built AI voice campaigns that ran across 28 live branches of a multi-location lending business, including a 10-branch pilot with 13,159 AI calls.
Where AI-first redesigns go wrong
- No baseline. If you did not measure the old process, you cannot prove the new one. Capture the three numbers before you touch anything.
- Designing around the demo. The demo record is clean. Design for the 30% that arrive incomplete.
- Deleting the step that was actually the control. Some "pointless" checks are the only reason a bad record never reached a customer. Ask what each step was catching before you remove it.
- Redesigning the process, not the job. If the freed hours are never reassigned, they silently refill with the same category of work and the gain disappears.
- One person holds it. A redesign that lives in one head is a risk, not an asset.
The one-week redesign sprint
A practical container for all of the above:
- Day 1 — Observe. Watch the work. Draw the current state with numbers.
- Day 2 — Baseline. Lock the three numbers you will compare against, and where they come from.
- Day 3 — Redesign. Run the five moves and write the before/after ledger. Decide the human gate.
- Day 4 — Build the smallest version. One path, structured outputs, logging from the first run.
- Day 5 — Run it on real records with a human approving everything. Compare against the baseline. Write down what surprised you.
Then repeat on the next workflow. The compounding comes from the habit, not the first build — which is the argument in Kaizen and continuous improvement.
Where to go next
AI-first workflow design is what an AI-native company does routinely and everyone else does once. To learn how the individual steps are built, read agentic workflows; to see why redesigning one process beats optimizing five, read systems thinking for AI automation. To practise the whole loop on real revenue work, start with Prove value fast.