AI Agents
AI agents · agentic AI
AI agents for $5M–$50M companies that still quote on spreadsheets.
An AI agent is software that takes a job from request to finished and asks a person only when a rule says so. For a company your size that’s a quote assembled overnight, an order keyed into the ERP before the customer calls, a schedule rebuilt when a machine goes down. We find which of your workflows should get one, price each in hours a year, and tell you which ones to leave alone.
94%
of mid-market companies use generative AI. 2% run it at scale.
Kaufman Rossin / NewtonX, Dec 2025, n=100, $5M+ revenue
95%
of enterprise AI pilots show no measurable P&L return.
MIT NANDA, Aug 2025, 300 deployments
40%+
of agentic AI projects will be canceled by the end of 2027.
Gartner, Jun 2025
The distance between the first number and the second is a diagnosis that didn’t happen. That’s the part we do.
What an agent is
Three things get sold as AI agents.
Gartner counts about 130 real agentic vendors among the thousands selling the word. Here’s how to tell them apart on a sales call.
Assistant
You ask. It answers. You do the work.
Drafts the email, summarizes the call, rewrites the SOP. Useful. Also where most companies stop, and why 94% “use AI” while 2% run it at scale.
Automation
Fixed trigger. Fixed steps. Same every time.
A form row lands, a CRM record gets created, a Slack message fires. Cheap and dependable until the input changes shape, then it breaks quietly.
Agent
Given a goal. Picks the steps. Stops when the rule says stop.
A quote request lands. The agent pulls the drawing, the part history, and the price book, builds the quote in your template, flags the 2 lines outside your margin floor, and hands it to the estimator for a 5-minute approval.
The agent is the only one of the three that changes a job description. So it’s the only one that needs a diagnosis before a build.
Where agents pay off first
The 7 workflows we look at before anything else.
These are the jobs that still run on spreadsheets and email at most $5M–$50M companies. Each line says what the agent does and what stays with your person.
→
Quoting and costing. The agent assembles the quote from the RFQ, drawing, and price book. Your estimator approves it.
→
Order entry and PR-to-PO. The agent reads the PO, keys the order into the ERP, and matches it to the quote. Your CSR handles the 1 in 10 that doesn't match.
→
Production and service scheduling. The agent rebuilds the schedule when a job slips or a machine goes down. Your scheduler signs off on the new sequence.
→
Weekly reporting. The agent pulls the numbers and builds the report your controller rebuilds by hand every Monday. Your controller reads it instead.
→
Customer status updates. The agent answers "where's my order" from the ERP and the carrier feed. Your CSR gets the escalations.
→
Intake, triage, and approvals. The agent routes the request, checks it against the rule, and approves inside the limit. Your manager sees only what's over it.
→
Invoicing, collections, and close. The agent drafts the invoice from the shipped order and chases the 30-day aging. Your bookkeeper handles disputes.
The rule that kills most agent projects before they start
Eliminate before you automate.
Most tasks on an agent wish list shouldn’t get one. An agent carries a build cost and a maintenance cost, and a task that returns under 20% of someone’s time rarely pays either back.
So that’s the rule we apply to every workflow: under 20% removable, we recommend deleting the task. Over 20%, we scope the agent. The Blueprint hands you both lists, and the second one is usually shorter.
Gartner’s 3 reasons agent projects get canceled are escalating costs, unclear business value, and inadequate risk controls. The 20% rule handles the first 2 before anything gets built. The governance section below handles the third.
Worked example
A quoting agent, step by step.
This is the workflow an independently run Tier-1 automotive plant picked as its first pilot after mapping its quote cycle: spreadsheet to email to ERP and back. The method is what’s on offer; delivery there is paused under a corporate spending freeze.
The arithmetic for your plant: 90 minutes per quote × 20 quotes a week × 50 weeks is 1,500 hours a year on one desk. The calculator runs your numbers in 4 questions, no signup.
01
A quote request lands
Email or portal. The agent reads the RFQ, pulls the drawing and the part history from where you keep them.
02
It prices the lines
Price book, current material cost, lead times from the ERP. Anything outside your margin floor gets a flag for the estimator.
03
It builds the quote in your template
Same document your customer already receives. Nothing changes on their end.
04
Your estimator reviews a queue
Approve, edit, or reject. The 90-minute quote becomes a 5-minute review, and the estimator’s judgment is still the thing that ships.
05
Every decision is logged
The inputs the agent used, the flags it raised, who approved it. That’s the audit trail your controller will ask for.
Reference builds
Three production systems, labeled honestly.
Two are agents. One is a pipeline, and it’s on the page because it shows what an agent would have been overkill for.
Agent · E-commerce
Recommendation agent across site and email
A growing e-commerce brand replaced generic suggestions with an agent that picks products per shopper in real time. +35% AOV, +60% email conversions, payback in under 90 days, no new traffic.
Agent · Sports media
Game-recap agent for a regional broadcaster
Speech-to-text, drafting, and voice synthesis chained behind an editor’s approval. Recap production went from about 5 hours to 20 minutes, ~70% lower cost (~$30K/yr), +45% daily listens.
Pipeline · Travel marketing
Video creative pipeline, fixed steps
Agency shoots replaced with a generation pipeline. 95% faster turnaround, up to 80% lower cost (~$50K/yr), +25% CTR. The steps never change, so this one didn’t need an agent.
These are the kinds of systems the Blueprint scopes. A named eMediaAI client case goes here after the first delivery.
Governance
An ungoverned agent is a liability wearing a productivity costume.
An assistant that’s wrong costs you a bad paragraph. An agent that’s wrong sends the quote. These 5 controls ship with every agent we scope, and they’re written into the Blueprint before anyone builds.
Audit trail
Every action logged with the inputs it used and the person who approved it.
Human gate
The agent stops at the rules you set: a dollar amount, a margin floor, a new customer, a part it hasn’t seen.
Error handling
What it does when the ERP times out, the drawing’s missing, or the price book has 2 entries for the same part. Written before launch, tested before handoff.
Data boundary
Which systems it can read, which it can write to, and what never leaves your network.
A named owner
Someone on your team, plus an SOP, a Loom walkthrough, and a governance doc. Done when your team runs it 30 days without us.
Your people
Your team will hear "agent" and think "replacement."
We decline engagements whose goal is eliminating jobs. It’s in writing on the engagement letter.
The Blueprint scores every workflow on the hours it returns to the person doing it, and the change plan names who owns the agent afterward. The estimator in the example above still owns quoting. She just stops retyping part numbers.
Most teams hear that in week 1 and relax. The ones that don’t are telling you something about the workflow, and we put that in the report too.
How you get there
One door in. The agent comes after the diagnosis.
1. AI Opportunity Blueprint™
10 days, $5,000 flat. We map the workflows, apply the 20% rule, price each candidate in hours a year, and hand you a sequenced roadmap with the agents worth building and the tasks to delete. A day on your floor if you’re within a day’s drive of Fort Wayne.
2. Build
Scoped from the Blueprint and quoted as one project fee. Your team, your vendor, or us. Handoff is an SOP, a Loom, a governance doc, and an ROI report.
3. Someone to own it
A Chief AI Officer on retainer for companies that have a roadmap and no one to run it. One monthly fee, quoted after the Blueprint, 3-month minimum.
If the Blueprint doesn’t identify at least 200 hours a year of removable work, each with a named mechanism and an implementation estimate, you don’t pay. You keep the document either way.
Questions owners ask before the call.
What does an AI agent cost to build?
It’s scoped from the Blueprint and quoted as one project fee. Publishing a number before the workflow is mapped would be a guess, and a guess is what Gartner’s 40% were built on. The Blueprint itself is $5,000, fixed.
Do we need to pick a platform first?
Pick it after. We don’t resell platforms and we don’t pick one before the process is mapped. The Blueprint tells you what the agent has to do, which is the spec you need before you can judge any vendor’s demo.
We're a division or plant of a larger company. Does this work for us?
Usually not, and we’ll tell you on the fit review. If a corporate parent decides which tools you can buy, the Blueprint produces a roadmap you can’t act on. The exception is a plant with its own P&L and purchasing authority for pilots under a set amount. Bring that number to the call.
What data does the agent touch?
Only the systems named in the data boundary, and only in the direction named: read, write, or neither. Nothing is built that moves data off your network without that line in the governance doc.
What happens when the agent is wrong?
It stops and routes to the named person, and the event is logged with the inputs it used. Every agent we scope has that path written before launch. An agent without one is the kind that gets canceled in year 2.
How long before an agent is running?
The Blueprint takes 10 days. A first agent on a scoped workflow is typically a matter of weeks, and the Blueprint gives you the estimate for yours. We tell you when the honest answer is longer.
We already bought Copilot and nothing happened. Why would this be different?
Because a license is an assistant waiting for a workflow. Dabbling is the risk. The Blueprint starts from the workflow, prices it, and only then asks which tool, if any, should run it.
Know which workflow gets an agent before you fund one.
Bring the one process that hurts. Within 24 hours of the call you’ll get a one-page note naming the 2 or 3 workflows where an agent would pay off first, whether or not you book the Blueprint.
