Flowgrammer

Iron Man Systems, Not Terminator Systems: How We Design Human-Led AI

Flowgrammer’s design principle: AI handles friction while people keep judgment, empathy, sensitive exceptions, and accountability.

— Craig Major

There are two ways to bring AI into a business. One removes the person from the work. The other makes the person more capable at it. We build the second kind, and the difference is not a slogan. It shows up as specific decisions we deliberately leave with a human, written into the workflow before anything is built.

Two kinds of system

Most AI conversations start with subtraction: which roles go away. That framing produces brittle systems, because the knowledge that made the work possible was never written down, and the person who could catch a bad output is no longer in the loop. We build the other kind. AI becomes the suit: it retrieves, prepares, applies agreed rules, coordinates steps, and drafts the next action. The person keeps judgment, empathy, sensitive exceptions, and accountability. Iron Man systems, not Terminator systems. It is a memorable line, but the reason we use it is practical. A system designed around a responsible person has somewhere to escalate, someone who notices when the output is wrong, and a name attached to the outcome.

The design rule we apply

Before a build starts, we split every step of the workflow into three columns and get agreement on the split.

Software handles reliably Software prepares, a person approves A person decides
Capturing an enquiry, deduplicating records, updating fields, moving files, scheduling, logging, notifying, reporting Classifying intent, summarising context, drafting a reply, suggesting a priority or route Pricing, credit, refusal, escalation, complaints, anything sensitive, and any case the rules did not anticipate
If a step cannot be placed in a column, the process is not understood well enough to automate yet. That is a useful outcome of the exercise, not a failure of it.

A worked example: inbound lead qualification

This is the workflow pattern behind our AI lead qualification work, described as a design rather than a client result. No outcome numbers are claimed here.

What the system does

  1. Captures the enquiry from the form, inbox, or phone transcript and creates one record, deduplicated against existing contacts.
  2. Assembles context: company, size, location, source, campaign, previous conversations, and anything already known about the account.
  3. Scores fit against written criteria the client approved, not a vendor's generic model.
  4. Classifies intent and urgency, and attaches a confidence level to that classification.
  5. Routes the enquiry to the right person or queue and drafts a first reply in the client's voice.
  6. Logs the whole path so cycle time, volume, and routing accuracy can be measured afterwards.

What we deliberately leave with a person

  • Disqualification. A system can rank and flag. Telling a prospective customer no is a human decision with commercial and reputational consequences.
  • Low-confidence classifications. Below the agreed threshold, the enquiry goes to a person with the context attached rather than being routed on a guess.
  • Anything that reads as distressed, sensitive, or unusual. These get a person immediately, by rule.
  • Sending the first reply in high-value or ambiguous cases. The draft saves the typing. The person owns the message.
  • Changing the scoring criteria. Criteria are a commercial decision, reviewed by the owner of the pipeline, not adjusted quietly by the system.

What has to be true for it to work

Written qualification criteria. One system of record. A named owner for the pipeline. An agreed definition of a qualified lead. Consent handling that satisfies PIPEDA and CASL for any message the system sends. Without those, the automation only makes an unclear process faster.

Why this is commercially better, not just kinder

  • Failure is visible. A reviewer catches wrong outputs before a customer does.
  • The knowledge gets written down. Splitting the work forces the rules out of people's heads and into the system, which is what makes the process improvable.
  • Adoption is easier. A team that keeps the judgment work uses the system. A team that expects to be replaced by it works around it.
  • Accountability survives. When a regulator, a client, or your own board asks who decided, there is an answer.
  • It scales in the right direction. Capacity goes up because people spend their hours on decisions and relationships instead of retyping and chasing.

Where the principle does not apply

Not every step deserves a human. Requiring approval on high-volume, low-risk, fully deterministic steps creates a queue and calls it governance. Human review belongs where being wrong is expensive, where the case is unclear, or where a person is affected by the decision. Everywhere else, get out of the way.

How to apply it this week

  1. Take the workflow your team complains about most and list its steps.
  2. Put every step in one of the three columns above, and mark the steps nobody can place.
  3. Estimate the current cost with the Automation Opportunity Calculator.
  4. Write down the decisions that will stay human, and what happens when the system is unsure.
  5. Then, and only then, look at tools.

If you would rather do that with us, the AI Success Audit produces the map, the baseline, and the recommendation. When several systems need one ongoing owner, that is Fractional CAO work.

Frequently asked questions What does human-led AI mean in practice?

It means splitting every workflow step into three groups before building: steps software handles reliably, steps software prepares for a person to approve, and decisions a person makes. Pricing, refusals, escalations, sensitive cases, and anything the rules did not anticipate stay with a person. Does keeping humans in the loop cancel out the efficiency gain?
Not when review is placed deliberately. Human review belongs where being wrong is expensive or the case is unclear. Requiring approval on high-volume, low-risk, deterministic steps just creates a queue. Which decisions should never be fully automated?
Pricing, credit, refusal and disqualification, complaint handling, escalation, and anything touching health, employment, or financial eligibility. A system can rank, flag, and draft. A person should decide and be accountable. What happens when the AI is not confident?
The workflow routes the case to a person with the assembled context attached, rather than acting on a low-confidence classification. The confidence threshold is agreed during design and measured after launch.

Sources

Related services

Continue with Flowgrammer