Chapter 4 of 7 · ~3 min
The people in the loop
Harbour Hotels asked for a discount of twenty per cent on sixty mid-range kettles. Policy permits fifteen without reference to a person. The run reached that point, computed a provisional quotation at fifteen, and stopped. Some forty seconds later the request reached Tom, with the company, the quantity, the discount requested and the discount permitted, what the research had established and what it had not, and two buttons. Tom asked the assistant provided alongside the request to summarise the research, read the summary, and approved the twenty. The quotation went out at twenty, the deal was recorded in the CRM at twenty, and the record shows Tom's name against the decision, beside the fifteen the rule had proposed. Had he not answered within a day, the request would have passed to Sam, the sales lead.
Participants, not users
Most of what needs to be said about accountability in such a process is contained in that account. A workflow is not a chatbot with a single participant. It has several, and they are not interchangeable: the buyer, the representative, the sales lead who receives escalations, and the agent, which is also a participant and has less authority than any of the people. Each acts under a role. Roles are declared in the workflow and are bound to individuals or teams when a run begins, so that no names appear in the process itself. A step that requires a person names the role and waits, whether for a minute or for a week, and the person answers from wherever he or she happens to be working. Nobody is obliged to keep a window open.
What the agent may do
The agent's role deserves particular attention, because it is the source of most questions about accountability. A model step acts under a role of its own, with exactly the access its step has been granted, never with a person's login and never reaching a system the workflow has not connected. People carry authority; the agent carries only the scope it has been given. When the record states that a step looked something up or drafted a reply, it also states which role did so and on whose authority, and the answer is never "the system". The decision to grant the buyer twenty per cent was Tom's, taken with a model's assistance and recorded as his.
Which decisions a person must own
For Nadia the practical consequence was a question of design rather than of technology: which decisions must be owned by a named person? A discount above policy, certainly. The declining of an enquiry from a country the company does not serve, on the other hand, is a matter for a rule, and the rule declines it courteously. Whether a buyer's follow-up question warrants a telephone call is a matter for a person, and the workflow passes it to one rather than guessing. The line between the two piles described in chapter 2 is also the line between what the agent may do on its own and where a person must be placed, and it is drawn by the owner of the process rather than by the model.
Experiment
Runs in your browserThe recorded run, replayed step by step in the inspector. Find the representative's step and read what the representative was shown before deciding. In this recording the representative's role was played by a test participant; the steps, timings and outputs are the run's own.
Run 64dbc007, a test run of the private staging deployment, driven by the coding agent as both participants. Twelve steps from the form to the seven-day wait; one minute of wall-clock time.
Agent · agent · research
research
- Input
- Instructions with the four ranges; the enquiry; existing customer: false. One tool: web_search. Output schema with five fields.
- Execution
- Call 1: 931 tokens, 3.4 s, requested web_search. The search tool returned an error. Call 2: 4,374 tokens, 18.3 s, final answer. Schema check passed first time.gemini/gemini-2.5-flash · 5000 in / 305 out · 21.8 s
- Output
- fit: "unknown" · reasons: "…could not be retrieved due to a failed web search…" · evidence: "Web search for Harbour Hotels in the United Kingdom failed. Based solely on the provided information: …" · suggested_products: "Birch" · questions_to_ask: one question about room aesthetics
- State write
research = "{fit: \"unknown\", reasons: …, suggested_products: \"Birch\", questions_to_ask: …, evidence: …}"
Two model calls in twelve steps, both bounded, neither touching a number. One decision by a person, recorded under their role next to what the model proposed. Every other step is a rule, a fixed call or a message, and the whole run can be replayed exactly as it happened.
Big question
Which decisions in your process must be owned by a named person, and where in the process do they sit at present?
