--- title: "2 things business leaders should do before building an AI agent" description: "Automating the wrong workflow only makes the problem faster. The Diagnose and Redesign stages of the FDO method, with every tool worked through on an inbound lead workflow." url: https://theclosedgap.com/writing/diagnose-then-redesign/ author: Madina Umbetova --- [Writing](https://theclosedgap.com/writing) The method · 13 September 2026 # 2 things business leaders should do before building an AI agent Diagnose, then redesign: the 2 stages I run before any build, with every tool on the page. 13 September 2026 9 min read [Read on Substack](https://theclosedgap.substack.com/p/2-things-business-leaders-should) There are around 80 lead agents for SMEs, and they cover agent jobs that support lead generation, qualification, conversion and growth. Not all of them are really agents. Some of them could be easily done as workflows. But important is that before you choose any agent, you should answer two questions. What exactly are you trying to improve, or which problem are you trying to solve? And second, are your company and your processes ready for agentic work? And this is what I’m helping you to do today. The process an agent would land in is usually older than the technology. It grew around people, systems, workarounds and past mistakes. An agent placed into that workflow inherits all of it, and it can make the existing process run faster, including the parts that no longer deserve to exist. > This is the sequence I use: > > **DIAGNOSE → REDESIGN → ARCHITECT → BUILD → DEPLOY → PROVE** > > Today we work only on the first 2 stages. **Diagnose** means to see how the work really happens today: what starts it, what information comes in, who does what, where the decisions sit, and where time or money leaks out. **Redesign** means to challenge that workflow before any technology enters it: which steps should disappear, which can be combined or moved earlier, which need explicit rules, and which must stay with a human. **That is the FDO approach: business problem → business job → solution.** The worked example through this article is the system I mapped: inbound lead research and qualification. Here is the situation it starts from. Leads drift in through various channels. It could be the LinkedIn inbox, emails, WhatsApp, or web form submissions, depending really which channels you advertise and which channels you personally use a lot. If you are the owner who does active sales, then mostly a lot will come to you: WhatsApp, SMS, or even calls. It’s very rare when the lead is immediately addressed, and it’s also very rare when the lead comes with hundred percent required information. Normally it’s very hard to understand immediately: is it a real lead? Is it your lead, or is it just a noise? And often that job requires additional work to complete: either callback, or send verifying questions, or send a questionnaire to fill, which is resource demanding for both your company and for your potential client. That is much more than a request for a lead-qualification agent. It is a workflow with evidence, decisions, actions and boundaries. ## Do it yourself Pick one workflow that is frequent enough to matter and specific enough to map. Open a blank document. Work through the tools in order. The output you are building: a problem definition, the actual workflow, a decision map, a leak diagnosis, a baseline, a redesigned flow and a work allocation. ### Step 1: Diagnose At the end of Diagnose you can show what happens today, name the decisions hidden inside the work, show where value leaks, and define what better performance would mean. ### Tool 1: the workflow diagnosis canvas Map the workflow as it happens in practice; the SOP usually shows less than that. Answer each of these for your workflow: - **Business outcome.** What result should this workflow create? - **Trigger.** What event starts it? - **Inputs.** What information exists at the start? - **Information gap.** What is unknown at arrival? - **Actual flow.** What happens, in order, from trigger to outcome? - **Actors.** Who performs or approves each part today? - **Systems.** Where does information enter, live and move? - **Handoffs.** Where does ownership change? - **Decisions.** Where does someone need to judge, choose or approve? - **Evidence.** What information supports each decision? - **Exceptions.** What breaks the usual path? - **Friction.** Where is there waiting, searching, repetition, rework or loss? - **Baseline.** What are the current volume, time, cost, quality or conversion measures? - **Better outcome.** What would materially better performance look like? For the lead example: the outcome is that every valid inbound lead is understood, handled appropriately and visible to the owner. The trigger is any incoming communication. The inputs may be as little as a name and a short request, and the person context, company context, intent and urgency may all be unknown at arrival. The baseline must be measured before any build: inbound volume, response time, research time, proportion handled, conversion by route, ageing leads. ### Tool 2: draw the actual workflow The map should show decisions, and decisions are where most maps go blank. The pattern: **TRIGGER → INFORMATION → DECISION → ACTION → NEXT DECISION → OUTCOME** ![The inbound lead workflow, mapped: any incoming communication, capture the facts, research person, company and context, decide whether there is enough evidence to qualify, decide priority and route, and every valid lead lands in the daily owner briefing.](https://theclosedgap.com/writing-assets/diagnose-then-redesign/diagnose.png) Maybe you already have enough leads, but nobody properly qualifies them. Maybe good prospects disappear because follow-up happens two days too late. ### Tool 3: the decision map “Qualify the lead” looks like one task. When I mapped it, it was 6 separate decisions, each with its own evidence, its own owner and its own cost of being wrong: 1. **Is this a genuine inbound enquiry with a usable record?** Evidence: the original message, channel, available identity. If wrong: medium. 2. **What can we establish about the person, company and context?** Evidence: permitted public sources. Known facts, inferences and unknowns stay separate. If wrong: medium. 3. **Does the lead meet the explicit qualification criteria?** Evidence: the evidence record plus owner-approved criteria. If wrong: high. 4. **What priority and route apply?** Evidence: the qualification result, intent, urgency, customer status. If wrong: high. 5. **What response is genuinely appropriate?** Evidence: the original request, the route, approved response boundaries. If wrong: high. 6. **Does a human need to intervene?** Evidence: ambiguity, value, sensitivity, contradictory evidence. If wrong: high. The point of this map is to make the operating boundary explicit before a system crosses it. ### Tool 4: the workflow leak test Where are you losing good prospects today? Go stage by stage. Where are leads getting lost because something is too slow, inconsistent, repetitive, dependent on someone remembering, or simply not happening? Mark each step in your workflow with the leak codes that apply. A step may have more than one. - **W · Wait.** Does work sit untouched? - **S · Search.** Does somebody repeatedly look for information that could arrive with the work? - **R · Repeat.** Is information entered, researched or explained more than once? - **H · Handoff.** Does ownership change without adding useful judgment? - **J · Judgment.** Are people applying implicit or inconsistent criteria? - **E · Exception.** Does an unusual case break the path or require improvisation? - **L · Loss.** Can a lead, request or decision disappear from view? - **RW · Rework.** Does someone redo work because the first pass was incomplete or unreliable? In the lead example: a message sitting unseen in an inbox is W and L. Manual company research is S and R. Different people assessing fit is J. The owner drafting every reply from scratch is W and R. A promising lead that is not ready is L and E. And at the end of the day, the test is L: can the owner see every lead, priority and unresolved question in one place? ### Tool 5: the Diagnose exit test Do not begin redesign until you can answer each of these with evidence. - What business outcome are we improving? - What actually happens today, including the informal work around the documented process? - Where are the consequential decisions? - What evidence does each decision require? - Who owns each decision, and who should be informed? - Where do wait, search, repeat, handoff, judgment, exception, loss and rework occur? - What is the baseline today? - What would materially better performance mean? - Which facts remain unknown and need measurement or an owner decision? For the lead workflow this means: the team can show the real inbound channels, the current response and research pattern, owner-approved qualification criteria, escalation conditions, response permissions, the nurture mechanism, and the daily briefing requirement. Any missing policy is a decision the owner makes before the system runs. ### Step 2: Redesign The difficult part is not finding an agent idea. It is knowing where to put it. You now have a map of the current workflow. The next task is to challenge every step, decision and handoff. The old sequence has no automatic right to survive. ### Tool 6: the redesign matrix Run each step through these levers: - **Eliminate.** Why does this step exist? Does it still create value? - **Simplify.** Can it require fewer actions, fields or decisions? - **Combine.** Can related work happen once, in one place? - **Standardise.** Can the criteria, required information or output format be made explicit? - **Resequence.** Should it happen earlier, later or in parallel? - **Change ownership.** Who or what is best placed to own it? - **Change information flow.** Can the right evidence arrive with the decision instead of being searched for later? - **Add a boundary.** What must stop for human review, clarification or approval? Applied to the lead workflow: the owner reading each message and starting from a blank page becomes a structured record captured at arrival. Separate manual searches become research that arrives with the record. Fit assessed from experience alone becomes explicit owner-approved criteria with the evidence shown. High-value or sensitive cases stop blending into the normal queue and get detected early, with a human brief created before any external action. Scattered updates become one record that feeds the daily briefing. ### Tool 7: the work allocation matrix Only after redesign do you decide who or what should own each part. The complexity I would only recommend based on the company real situation. The readiness of your team, the readiness of your internal processes, the readiness of your data, availability of that data, and how much control you would like still to have. Agent, in this particular case, is the most advanced version. Do you really need it? Not necessarily. The more advanced feature, the more potential problems you might get. Automation, or workflow automation, I think is very robust. There is only one possible option: if x then y, and it works. So I would first suggest to understand your situation, your real needs, and then to decide what exactly you do. > Use the lightest system capable of fixing the problem: Chat → Automation → Workflow → Agent. - **Remove it** when the step creates no useful outcome. Stop copying the same contact details across disconnected notes. - **Deterministic automation** when inputs, rules and outcomes are stable and approved. Capture the fields, create the record, apply routing rules, assemble the daily summary. - **AI assistance** when synthesis helps but a human owns the decision. Research public context, summarise evidence, draft a recommendation. - **Agent-owned work** when bounded autonomy creates value and mistakes are recoverable. Assemble context, maintain the evidence record, act within explicit, approved boundaries. - **Human-owned work** when risk, ambiguity or commercial consequence is high. High-value opportunities, sensitive cases, contradictory evidence, policy exceptions. ### The before and after ![Redesign the workflow before you add AI: the same lead process before (seven manual, person-dependent steps) and after (capture, enrich, assess, decide, act and track).](https://theclosedgap.com/writing-assets/diagnose-then-redesign/redesign.png) > In this redesigned flow, no agent exists yet. What exists is the decision an agent would have to serve: the research and record work is agent-shaped, the sensitive calls stay human, and the next stage, Architect, is where that gets built. You can get a very easy walk through these 2 stages in my [free template pack](https://theclosedgap.com/resources/workflow-diagnosis-redesign-kit). **Subscribe to get it.** Start with one workflow, map what actually happens, find the decisions and the leaks, and redesign the logic. Then decide what deserves automation, AI assistance, agent autonomy and human attention. Don’t automate the stage with the most work. Fix the stage with the biggest business loss. Before you spend anything on building AI agents, this is the job you can easily do as the owner or business leader, and it normally costs you nothing. And with that, you will be absolutely ready with the decision: which agent do I really need? And the second: your systems will be ready for that agent to deliver measurable business impact, not just an automated mess. The Closed Gap ## One method _every week._ Written by an operator running a $0.8bn global portfolio, for operators. Subscribe free and the next one lands in your inbox. The Closed Gap on Substack. No spam, unsubscribe in one click.