Our approach

“Can AI replace it?” is the wrong question. Almost.

Nobody replaces a process. You replace steps. Some belong to AI, some to plain automation, and some should stay exactly where they are: with a person. Our job is to tell them apart, honestly.

01 · What we believe

Four principles behind every verdict.

Steps, not processes.

A process is never one thing. Split it, and the yes/no question turns into four useful answers.

Rules before models.

If a rule can do it, a rule should. AI is for the messy parts: reading, drafting, interpreting.

A person owns the outcome.

Commitments, exceptions, and relationships keep a human owner. Always. That is a design choice, not a limitation.

Say what we don’t know.

We assess from your description, not your systems. Every report lists its assumptions and open questions.

02 · The spectrum

Every step lands in one of four places.

We ask the same questions of every step, in the same order. The answers push it left or right.

AI candidate

Let AI do the work.

Understand, extract, or draft within a clearly defined task with verifiable output.

We look for: unstructured input, repeatable intent, a way to check the result.
Human review

A copilot, with checks.

AI prepares the work. A person reviews it and owns the decision before anything leaves the building.

We look for: customer-facing output, meaningful cost of error, judgement about tone or context.
Standard automation

Sometimes, rules win.

Move data, trigger reminders, connect systems. Deterministic, cheap, and easy to trust.

We look for: structured data, stable rules, existing integrations or APIs.
Keep human

Keep the human touch.

Negotiation, sensitive exceptions, and relationships need an accountable owner.

We look for: commitments, conflict, ambiguity that a model cannot resolve safely.
03 · How a step gets its verdict

Five questions, asked of every step.

Not a scorecard, a conversation. The point is to make the reasoning visible so you can disagree with it.

  1. 01
    What comes in, and what comes out?Structured records point to rules. Emails, calls, and documents point to AI. Both point to a clear interface.Structured → Automation · Unstructured → AI
  2. 02
    How often is this step the same?Repeatability is what makes any of this worth building. Rare exceptions can be routed; frequent ones mean the step is really a decision.Rarely varies → left · Often varies → right
  3. 03
    What happens when it goes wrong?A typo in a spreadsheet and a wrong price in a quote are different problems. Consequence decides how much review the step needs.High consequence → Human review or Keep human
  4. 04
    Can the result be checked?AI output you can validate, against a schema, a source record, a rule, is safe to use. Output you can only trust is not.Verifiable → AI · Unverifiable → Human
  5. 05
    Does it require a commitment?Promises to customers, money leaving the company, decisions about people. These stay with someone who can be asked why.Commitment → Keep human, always
04 · What you get

A technical brief, not a sales pitch.

See an example report →
VerdictOne line you can act on.

What the process mostly is, and the single best first move.

Step by stepEvery step, with reasoning.

The verdict, the proposed implementation, and the human control that goes with it.

Technical approachA stack we would actually build.

Tools, architecture, what to verify first, and when not to use an agent framework.

AssumptionsWhat we took for granted.

Open questions and assumptions to validate with the people who own the process.

05 · Honest limits

What an audit can’t tell you.

We work from your description. That is enough to point in the right direction, and not enough to promise results. So we don’t.

  • It has not seen your systems. Integrations, permissions, and data quality are assumptions until someone checks them.
  • It is not a savings estimate. We show the workload you told us about. What you save depends on a pilot, not a page.
  • It is a first assessment. More context changes verdicts. That is why every report ends with “reassess”.
  • It will sometimes say no. If the honest answer is “don’t automate this”, that is the answer you get.
Who’s behind it

An initiative by Arcgent.

Can AI replace it? started as the question every B2B team kept asking us. We build the workflows we recommend: n8n, bounded AI steps, human approval gates. Which is exactly why we’d rather tell you when you don’t need one.

Book a discovery call with Robin
Ready when you are

Put one process through the audit.

Three questions, a couple of minutes, and a step-by-step verdict you can share with your team.

Audit my process