AI Implementation · 9 min read
Agents Fit
The Omni model: one managed AI Employee owns one recurring workflow; specialist employees add capacity around the same business context; your team keeps the judgment calls.
A clinic manager calls us because her front-desk team spends two hours every morning answering the same ten questions: insurance accepted, hours, parking, what to bring to a first appointment, prescription refill timing. The questions are predictable. The answers are documented. But the volume is high enough that the team starts the day tired and is slower with the patients who actually need attention.
That is the canonical "agent fit" scenario. The workflow is high-volume, low-variance, and answerable from a known knowledge source. An AI agent can handle it. The trick is identifying which of the remaining 30 workflows that clinic runs in a year are the same shape, and which are not.
"Agent fit" is the term we use at Omni Studio for that identification work. This article walks through how we assess it, how we structure the workflows that pass, and where we deliberately step back and recommend a different solution.
What "agent fit" actually means
An AI agent is a software process that reads input, reasons over a defined knowledge set or tool set, and produces output or takes a defined action. That definition matters because it constrains where agents work well.
An agent is a fit for a workflow when three conditions hold:
- The input is structured or semi-structured. Free-text is fine, but the underlying request must be classifiable into a bounded set of intents.
- The correct output is determinable from a known source. A policy doc, a CRM record, a pricing matrix, a scheduling system. Not from a human's judgment in a novel situation.
- The cost of a wrong answer is bounded and recoverable. A wrong FAQ answer is a mild embarrassment. A wrong contract interpretation is a liability. These are not the same category.
If all three hold, the workflow is a candidate. If any one fails, the workflow is not an agent job — it is either a human job, a hybrid job, or a process redesign before it becomes an agent job.
This framing is consistent with how McKinsey has been tracking AI deployment maturity: the value comes from re-engineering specific workflows, not from sprinkling AI across the org chart. Their 2024 state-of-AI work found that companies pulling ahead are the ones redesigning processes around what the technology actually does well, rather than retrofitting it onto existing chaos.
The three questions we ask before any deployment
Before we build anything, we run the workflow through three questions. They are unglamorous. They are also the difference between an agent that pays for itself in a quarter and one that gets killed in a steering-committee meeting.
Question 1: How often does the right answer change?
If the answer set is static — store hours, onboarding steps, return policy — the agent is straightforward. If the right answer depends on who is asking, what they bought, what their contract says, and what their account history looks like, the agent needs to be wired into systems and given clear reasoning rules. That is still doable. It just moves the project from a one-week build to a four-to-six-week build, and from a single prompt to a tool-using agent with retrieval and writes back to the source of truth.
Question 2: Where does the human need to review?
Every agent we ship has at least one human review point. The question is where it lives. For an inbound sales qualification agent, the review point might be: the agent qualifies and books the meeting, but a human reviews the meeting brief 60 seconds before the call. For a support agent, the review point might be: the agent drafts the response and queues it for a human to send on any ticket tagged "billing dispute" or where the customer's sentiment score crosses a threshold.
This is the approval-gated automation pattern. It is not a workaround for a weak agent. It is the design. The agent does the repetitive work; the human owns the consequential decision.
Question 3: What is the failure mode?
If the agent hallucinates an answer to a patient's insurance question, the clinic handles one annoyed phone call. If the agent hallucinates an answer to a contractor's lien waiver question, the contractor may have a legal exposure. Same shape of workflow, very different blast radius. We size the agent's autonomy to the blast radius. Higher blast radius means tighter constraints, more retrieval grounding, and a shorter leash before a human sees the output.
Anatomy of a well-fit workflow: inbound support triage for a property management firm
Concrete example. A property management firm in the Southeast runs 1,400 doors. Their after-hours inbox gets 80 to 120 messages a night: lockouts, noise complaints, maintenance requests, lease questions, and a long tail of "my rent didn't go through, what do I do." Before us, a leasing agent was on call from 6 p.m. to 7 a.m. and was burning out.
The workflow we built:
- Intake. Messages arrive via email, web form, and SMS. A router classifies each into one of nine intents using a combination of keyword signals and a small classifier model trained on six months of historical tickets.
- Retrieval. For each intent, the agent retrieves the relevant policy from the operations binder we converted into a structured knowledge base. Lease questions pull from the lease template. Maintenance requests pull from the vendor list and the unit's maintenance history.
- Draft and route. The agent drafts a response or, for maintenance requests, opens a work order in the existing system. Every draft is held for human review unless the intent is in a pre-approved list (lockout protocol, rent confirmation, package acceptance notice).
- Human review. The on-call leasing agent reviews the queue once at 7 a.m. Approves the drafts, edits a handful, and the system sends. Anything flagged urgent (active water leak, lockout, safety issue) is escalated in real time via SMS to the on-call phone.
- Logging. Every interaction is logged with the intent, the retrieved source, the draft, and the human edit. We review the edit patterns monthly and tighten the agent's retrieval or escalate list accordingly.
The on-call agent now sleeps through the night unless there is an actual emergency. The response time to tenants dropped from an average of nine hours to eleven minutes during after-hours. The leasing agent's retention is no longer an emergency project.
None of that required replacing the leasing agent. It required giving her a triage layer that handles the repetitive 70% and surfaces the consequential 30% to her, already organized.
Where agents do not fit
We turn down work sometimes. The honest list of what is not an agent fit, in our experience:
- Novel negotiation. First-pass commercial deals, complex renewals, partnership terms. These need a human who can read the room, make concessions, and trade variables. An agent can prepare the brief. It should not be in the room.
- High-stakes single-shot decisions with irreversible outcomes. Medical diagnosis, legal advice, certain underwriting calls. The blast radius is unbounded, and the agent's reasoning is not inspectable enough yet for these settings.
- Workflows with no source of truth. If the right answer lives in three different people's heads and they disagree, an agent will pick one of them at random and call it policy. Fix the source of truth first; the agent comes second.
- Anything where the volume is too low to learn from. If a workflow runs ten times a year, the data you need to evaluate the agent's performance does not exist. Human is cheaper.
This list overlaps with what Harvard Business Review has been writing about the limits of generative AI in the enterprise: the technology is a powerful pattern-completer, but pattern-completing on a five-sample-per-year workflow is just guessing with a confident tone. The honest assessment of where it does not belong is part of the deliverable, not a footnote.
How we roll out: the three-week pattern
For workflows that pass the three-question test, our deployment pattern runs roughly three weeks, with deliberate gates:
- Week 1, mapping. We sit with the operations lead and walk through the workflow step by step. We identify the inputs, the decision points, the existing systems, and the failure modes. Output: a one-page workflow map with agent-fit annotations on each step.
- Week 2, build and shadow. We build the agent, wire it into the systems, and run it in shadow mode. It produces drafts and takes actions in a test environment, but nothing reaches a customer. We compare its outputs to what a human would have done on the same inputs and tune.
- Week 3, gated live. The agent goes live, but every consequential output still hits a human review queue. We watch the edit rate. Once the edit rate drops below a threshold we set with the client (typically 15–20% of drafts require substantive human change), we begin auto-approving the pre-cleared intent categories.
Throughout, we hold to the NIST AI Risk Management Framework's posture on human-AI configuration: meaningful human control, transparency about what the agent is doing, and the ability to override or shut down any action in under sixty seconds. That is not regulatory overhead. It is what makes the rollout survivable when something unexpected happens, which it will.
Frequently asked questions
How do I know if my workflow is a candidate without hiring a consultant?
Ask the three questions above. If you can answer "yes" to all three and you have at least one person on your team who can describe the workflow step by step, you have a candidate. If the workflow lives in scattered Slack messages and tribal knowledge, you have a documentation project, not an agent project yet.
Will an agent replace my team's roles?
No. In every deployment we have done, the agent handles the repetitive intake, drafting, and routing work. The human role changes: less inbox time, more time on the cases that need judgment, relationship work, and exceptions. Teams we work with typically see their people stop quitting from burnout before they see any headcount change.
What does it cost to find out if a workflow is a fit?
The mapping week is the cheapest part of the engagement and the highest-use. It produces a written assessment of every workflow you flag, with a fit score and a recommended next step for each. We run it as a fixed-fee audit because the output is more useful to you than a sales deck.
How long until the agent pays for itself?
We do not promise payback windows and you should be skeptical of anyone who does. What we will say: for the workflows we typically take on — inbound triage, lead qualification, scheduling, document drafting — the time saved per week is large enough that the build cost recovers within one to two quarters, based on the labor cost it displaces. Your numbers will depend on your fully-loaded hourly cost and your volume.
What happens when the agent is wrong?
The human review point catches it. If a category starts producing bad drafts at a higher rate than expected, we have kill switches that pause the agent in under sixty seconds and route everything to humans. We review every flagged interaction in a weekly tuning session and either tighten the agent's constraints or move that intent category back to human-only.
Where to start
If you are reading this and thinking about two or three workflows that match the shape described above, the practical next step is a one-week mapping engagement. We work with your operations lead to walk through the candidate workflows, score each on fit, and produce a written recommendation with rough build cost and a deployment order.
Book a free AI automation audit and we will send back a written assessment of your top candidates within a week. No pitch deck, no obligation to build.
