Most lists of agentic workflow examples describe what an agent could do. These nine ran in real online businesses: a pre-call brief, enquiry scoring, a welcome sequence, small proposals, a daily ad recommendation, cancellation saves, a community assistant, sponsorship replies and launch sends. Each has its numbers, and five sit beside the same idea built wrong.
The businesses are ones I’ve worked with. They’re anonymous by agreement, so I describe each only by its type and annual revenue. Every figure came from the business itself, and where a result has a competing explanation I give that too.
The nine wins come from businesses at $185K to $1.1M a year. Several of the failures come from bigger ones, up to $2.4M.
Read all nine and a pattern shows up. The result usually came from a check, a permission, a written file or a decision about who decides, and less often from the quality of the model. The counts are after the examples.
Contents
- The nine at a glance
- What makes a workflow agentic
- 1. A brief before every sales call
- 2. Enquiries scored against written criteria
- 3. A welcome sequence assembled from emails you already wrote
- 4. Proposals that stop waiting on the founder
- 5. A daily ad recommendation with no access to the budget
- 6. Cancellation saves that offer a change before money
- 7. A community assistant that says what it is
- 8. Sponsorship replies with four terms reserved for the founder
- 9. Launch sends where the agent can add people and never remove them
- What the nine have in common
- You may not need an agent at all
- How to pick your first one
- Questions people ask
The nine at a glance
| Workflow | Business | What changed | Caveat on the result |
|---|---|---|---|
| Pre-call brief | $290K solo consultant | Show rate 62% to 81% across 94 calls | The brief and the confirmation changed together |
| Enquiry scoring | $230K independent consultant | Average project value $6,200 to $11,400 | None on this figure |
| Welcome sequence | $185K course creator | 30-day purchase rate 2.4% to 5.1% across 2,140 subscribers | Before and after, with no split test |
| Proposals under $5K | $460K design studio, three people | Turnaround four days to under three hours | None on turnaround. The close-rate lift rests on four weeks |
| Daily ad recommendation | $690K coaching business | Acquisition cost $412 to $317 while spend rose 19% | Cost moved for several reasons, including a July landing page change |
| Cancellation saves | $890K coffee subscription | Save rate 11% to 27% across 412 requests | None on the save rate |
| Labeled community assistant | $340K membership | Median reply time 7 hours to under 2 minutes | None on reply time. Churn also fell, for several reasons |
| Sponsorship replies | $1.1M newsletter | Median reply time 3.2 days to 6 hours | None on reply time. Revenue per issue rose while the list grew |
| Launch sends | $950K course business | A rejected removal list produced 34 sales, $42,500 | Whether those buyers would have bought another way is unknown |
What makes a workflow agentic
An agentic workflow has a decision in the middle that software makes. A form that triggers an email is automation. A system that reads an enquiry, scores it against your criteria and drafts a yes or a no is agentic, because it judged something.
In the examples below the agent’s own decisions are narrow. Automation, files and people do the rest. If you aren’t sure which a job needs, I compare the two in AI agents vs automation, and the design side is in my guide to agentic workflows.
1. A brief before every sales call
A solo consultant with $290K in annual revenue has an agent prepare her for every discovery call. It fires 24 hours before the call. Firing at booking would leave the research stale.
The agent reads two files of hers. One is a client profile with her fit criteria and three exclusions. The other holds 40 questions she collected over four years, tagged by situation.
It researches the prospect’s revenue model, last three public moves and current positioning, with source links, and writes “not found” where it finds nothing. Then it picks three questions from her file and drafts a two-line confirmation that mentions one specific finding. The confirmation sits in a draft queue and can’t send without her click.
Across 94 booked calls between May and July 2026, show rate was 81%, against a 62% baseline. Close rate on calls held went from 29% to 38%. The brief and the confirmation changed at the same time, so she can’t say which did it. She thinks the two-line confirmation matters more and has no evidence for that.
She built it herself in two afternoons, and the question file already existed in a notes app. She reads each brief for about 40 seconds and has never changed one. The agent never touches the call, and she has ruled out unreviewed sending for good.
2. Enquiries scored against written criteria
An independent consultant at $230K a year has an agent score every enquiry against five written criteria: budget signal, timeline, whether the decision-maker is present, scope fit, and a list of three industries she won’t work in. The excluded industries are checked first, and a match goes straight to a decline whatever the score. Where evidence is missing, the agent says so and doesn’t guess.
It drafts either a call invitation or a warm decline with a referral to one of four peers. She approves each in under a minute, and every disagreement goes in a log.
Of 83 enquiries in five months, 29 were invited and 54 declined. Before, she declined about one in five. Average project value went from $6,200 to $11,400, and annualized revenue rose to $312,000 on fewer projects. She overrode the agent 7 times, five of them to invite someone it had declined.
The hard part was writing the five criteria, which took two sittings. The build took about four days.
One risk to know about comes from a different business. A $1.1M fractional CFO firm added agents for qualification, call prep and proposal drafting.
Qualified opportunities went from 9 a month to 21. Close rate fell from 44% to 23%, and the median wait between a signed proposal and the first session went from 9 days to 26. Headcount didn’t change.
The firm blames the wait for the lower close rate. A bigger funnel with weaker leads in it is the competing explanation, and both are probably partly true. Check where your real limit is before you speed up the front. The full build is in my lead qualification workflow, step by step.
3. A welcome sequence assembled from emails you already wrote
A course creator at $185K a year, with 11,800 subscribers, used to send one welcome sequence to everyone. Now an agent builds a five-email sequence for each new subscriber from a library of 23 emails she wrote and tagged by situation, objection and stage.
Thirty minutes after signup, the agent matches the signup source and a survey answer to those tags, picks five emails and puts them in order. It hands back five email IDs. It has no authority to write a sentence. Her email platform records the survey tag and does the sending.
One check guards the output. If the agent returns an ID that isn’t in the library, the assembly fails and the subscriber gets the default sequence. That has happened 11 times.
Across 2,140 new subscribers between June and August 2026, the share who bought her $397 program in their first 30 days went from 2.4% to 5.1%. First-month unsubscribes fell from 6.8% to 4.1%. This is a before-and-after comparison with different traffic sources in each period, and she says plainly that she doesn’t know how much of the lift is seasonal.
The build took a weekend of tagging and about two hours of setup, with no developer. Every word a subscriber reads was approved before the system existed, so you can audit the whole thing by reading 23 emails. For the trigger, branch and exit rules behind sequences like this, see marketing automation workflows.
4. Proposals that stop waiting on the founder
At a three-person design studio with $460K in annual revenue, every proposal waited on the founder for a median of four days. Before giving an agent any authority, the studio ran it beside him.
For 34 proposals over about seven weeks, the agent wrote its version and the founder wrote his. The two matched on 31. All three misses were projects above $5,000 with custom scope.
So the line went there. Since early September 2026 the agent sends proposals under $5,000 without review, which is 71% of volume. Anything at or above $5,000 is blocked from sending and goes to his queue.
Turnaround on the small tier went from four days to under three hours. Close rate on that tier went from 41% to 56%, a figure that rests on four weeks of data in a month that has run strong for this studio before.
Prices live in one file the founder maintains, and it is the only source of any number. A proposal containing a figure that isn’t in that file fails before it can send.
A $2.4M productized SEO service shows what happens without that file. Its quoting agent learned its style from twelve past proposals. Two of them carried a 15% first-month discount that a former salesperson had offered without authorization. The agent copied it into 19 of 63 quotes.
Eleven of those closed, for $6,845 in discounts nobody approved. Four of the customers have since asked whether the discount repeats at renewal, and the business has no written answer.
An agent learns as much from your examples as from your instructions, so read every example before you hand it over. I cover the side-by-side method in how to turn an SOP into an AI agent.
5. A daily ad recommendation with no access to the budget
An online coaching business with $690K in annual revenue spends $31K a month on ads. Its agent has no spending authority. At 6am each day it produces a one-page recommendation: what to cut, what to scale and what to leave. Each ad set’s return over 3, 7 and 14 days sits side by side, with the booking lag noted, meaning how long buyers usually take to book.
The founder spends about ten minutes approving or rejecting. Two rules sit outside the agent. A cut recommendation on any ad set with less than 14 days of data is blocked. Any change above 20% of daily budget is flagged for explicit approval.
The agent also reads a one-paragraph file that states the measured booking lag and the minimum evaluation window. The founder wrote it in one afternoon with her media buyer. If you run your own ads, that paragraph is yours to write.
Over five months, blended customer acquisition cost fell from $412 to $317 while spend rose 19%. She approved 71% of the recommendations and reversed four decisions she had approved. The cost moved for several reasons, including a landing page change in July, and she doesn’t claim the agent is the only cause.
A $1.7M online fitness coaching business, spending $94K a month, gave its agent live spending authority. The agent rebalanced the budget daily on three-day return. Most of that business’s buyers book a call before purchasing, and the median lag is 9 days.
In August the agent made 41 reallocations. Twelve of them cut or zeroed an ad set that was unprofitable at three days and profitable at fourteen. Two had been July’s top performers and were never restarted. The business estimates $31,000 to $44,000 in lost August revenue, a modeled figure, and says August seasonality is a competing explanation.
The lag was known. It was in the founder’s head and in the CRM, and nowhere the agent could read.
6. Cancellation saves that offer a change before money
A specialty coffee subscription with $890K in annual revenue and 3,100 active subscribers received 412 cancellation requests between May and September 2026. An agent intercepts each one and sorts the stated reason into one of seven codes. “Unclear” is a valid code, and it routes to a person.
The agent must offer a change before it may offer money: a different roast, a longer interval, a paused month or a smaller bag. A discount exists for one reason only, price, and at one level only. Anything above that level is blocked.
Save rate went from 11% to 27%, which is 66 more subscribers kept. Nobody complained about the intervention. In the previous year, when people improvised the saves, 14 did.
The piece I’d copy first is the hard stop. If a subscriber asks to cancel a second time, the cancellation completes at once and the agent makes no further offer. It works because it is a permission and involves no judgment.
The file listing which offer goes with which reason took the founder a morning to write. It was the first time she had decided what she would and wouldn’t pay to keep a subscriber.
7. A community assistant that says what it is
A membership with $340K in annual revenue, 980 members at $29 a month, runs a community assistant on its own labeled account. It has its own avatar and a profile line saying it is automated and that the team sees everything.
It answers logistics questions and finds past sessions and resources. Anything about strategy goes to the founder, and the member gets a holding reply with an expected wait. The assistant can’t post as her, because it doesn’t have her login. The community manager owns the routing rules and reads every exchange each week.
Across 1,460 member questions between June and September 2026, median reply time went from 7 hours to under 2 minutes. Two members commented on the assistant, both positively. The founder’s forum time dropped from about eight hours a week to three.
Monthly churn fell from 4.6% to 3.1% over the period. It moved for several reasons, including a new onboarding call for members that launched in June, and the founder says she can’t separate them.
The build took about two weeks. Most of that went into an index of three years of session recordings, which is work a VA can do.
A $780K membership with 1,900 members did the opposite. Its agent answered forum questions under the founder’s name and photo, through her own login. Over about eleven weeks, 14 members said publicly that the replies didn’t sound like her. Four said “AI” outright, and three cancelled and cited it.
One answer contradicted a strategy she had publicly reversed in May. The agent was reading a knowledge base last updated in March. The business has no read on how many members noticed and said nothing.
If you run a community alone, copy the label and the routing, and never give an agent your own login.
8. Sponsorship replies with four terms reserved for the founder
A newsletter with $1.1M in annual revenue and 64,000 subscribers wrote one table before pointing an agent at sponsorship enquiries. Four kinds of term are reserved for the founder: exclusivity, performance guarantees, usage rights and audience access.
Any enquiry that touches one of the four blocks the automatic reply. The sponsor gets a holding reply, the founder gets a summary, and the agent can’t write terms language at all. Standard enquiries get a reply with a booking link. A part-time sales lead owns the rate card, the list of standard prices, and reviews ten standard replies a month.
Across 148 enquiries in five months, the agent handled 112 from start to finish and routed 36. Median reply time went from 3.2 days to 6 hours.
Sponsorship revenue per issue rose from $3,100 to $4,750. The business credits most of that to reply speed, and acknowledges the more obvious explanation: the list grew from 51,000 to 64,000 over the same months.
The table took the founder about an hour to write. In July a request for live event sponsorship fell outside both the table and the rate card. Automatic replies paused, as designed, and the rate card was updated in three days.
A smaller newsletter, with 31,400 subscribers and roughly $18K a month in sponsorship, ran the same job with no such table. Of 19 enquiries, six carried terms outside the rate card: two 90-day exclusivity requests, two performance guarantees, one request for perpetual usage rights and one list swap.
The agent flagged all six. Two of its draft replies still read like a soft commitment. Nothing went out, because the creator reviews every reply by hand. Nobody had written that review down as a rule, and no one owned nonstandard terms.
9. Launch sends where the agent can add people and never remove them
A course business with $950K in annual revenue ran a six-day launch in September 2026 with an agent managing send timing and segments. The agent can add a recipient to any send. It can never remove one, because its access to the list is add-only. That was a configuration choice and needed no development.
When the agent thinks someone shouldn’t get an email, it proposes a removal list with a reason for each contact. The launch owner approves or rejects the whole list. Any list more than 24 hours old blocks the send. If you run launches alone, that owner is you.
It proposed seven lists covering 3,890 contacts. The owner approved six and rejected one, which held 1,110 people tagged from a 2024 webinar. Those people went on to produce 34 sales worth $42,500.
That figure is exact, because the 34 buyers are identifiable. Whether they would have bought some other way is unknown. The launch did $412,000 against a $360,000 forecast, with no wrong-segment sends across 11 emails.
A $2.1M course business ran a five-day launch the same month with an agent that could remove people. It built its “already purchased” list from a customer export last refreshed before the launch opened. The export included 1,340 people who had bought a different, cheaper product.
They were dropped from the final two cart-close emails, which had carried 36% to 41% of revenue in the previous four launches. The launch came in $52,600 under forecast. The owner estimates $19,000 to $26,000 of that came from the removal, a modeled figure, and thinks the audience tiring of the price point was the bigger cause.
The first business named the trade-off before launch. Emailing someone who already bought costs an apology. Skipping someone who was about to buy costs the sale and leaves no trace.
What the nine have in common
These examples come from a set of 50 cases I keep from businesses I’ve worked with: 26 wins, 18 failures that cost the business something, and 6 problems caught before a customer saw anything.
In 24 of the 26 wins, the result came from a check that passes or blocks, a permission, a written file or a decision about who decides. In the other two it came from the quality of the agent’s method.
The owners follow a similar split. Twenty-five of the 26 wins had a named owner before launch. Fourteen of the 18 costly failures had no recorded owner when they failed.
You can see both in the nine. The second-ask stop, the add-only list, the 14-day block and the four reserved terms are checks and permissions. The question file, the 23 emails and the booking-lag paragraph are files. Each build also has a person who owns the decision, such as the consultant in the first example and the launch owner in the last.
The rule I use: a judgment goes to the agent. A fact goes in a file the agent reads every time it runs. A limit goes in a step that passes or blocks and has no opinion. Most of the failures above are a fact or a limit that was never written anywhere the system could read or enforce.
You may not need an agent at all
Common advice says to start with an agent. The solo operators in my case set point the other way. One group of eight cases is all solo operators at $250K a year or less who built their own systems, and seven of the eight use no agent.
One is an executive coach at $185K a year who works alone, with 14 clients on monthly retainer. After each session she runs one saved prompt against a folder for that client. The folder holds their goals, every prior session note and what they agreed to do. The prompt writes a summary for the client, a private note for her and three check-in prompts.
Across 112 sessions in six months, her admin after each session went from roughly 40 minutes to 9. She took on four more clients without a longer week. It took two evenings to build. She edits every summary before it goes out and changes something in about one in four.
If the job has no decision in the middle, or you can make that decision yourself in ten seconds, a prompt you run by hand is the right build. I go through that choice in when to use AI agents. For the steps that should stay with a person even when you do automate, see when not to automate.
How to pick your first one
Pick the job where you already make the same decision again and again, and where you could write your rule down. If you sell on calls, start with the pre-call brief. It took two afternoons, it reads two files, and it can’t send anything on its own.
Whichever you pick, build in this order.
- Write the file first. The sponsorship table took an hour and the retention offers took a morning. The design studio’s founder spent most of three part-time weeks writing down which projects had to come to him, rules he had been applying by instinct.
- Decide what the agent must never do, then remove its ability to do it. The consultant’s confirmation can’t send without her click, and the launch agent has no way to remove a contact. An instruction is weaker: in three of the cases that went wrong, someone had already tried to fix the problem with an instruction, and it didn’t work.
- Run it beside yourself before it acts alone. The design studio compared 34 proposals before the agent sent one.
- Name the owner, even when it’s you, and one number that would tell you it worked. Record that number for the period before you switch the agent on. I explain how to choose it in how to tell if your AI agent is working.
- Count how many times it will run in a month and what each run costs. The token calculator prices a workflow of up to ten steps.
I run a briefing agent of my own before every meeting that matters. The full build is in a Bionic Business issue, with its four passes and every prompt: Run This Briefing Agent. There is a matching issue for the lead qualification agent.
Open a blank file today and write the rule for the job you picked. If you can’t write it down, an agent can’t follow it, and finding that out costs you an hour.
Questions people ask
What is an agentic workflow?
An agentic workflow is a business process in which software makes at least one judgment in the middle: it reads something, decides, and then acts or hands off to a person. Fixed if-this-then-that automation makes no judgment. In the nine examples here, each judgment is narrow, and rules, files and human approval do the rest.
How is an agentic workflow different from automation?
Automation follows fixed rules: if this happens, do that. An agentic workflow adds a step where software reads something and judges it, such as scoring an enquiry or sorting a cancellation reason. In the examples here, fixed rules still surround every judgment.
Do I need a developer to build one?
Often not. The consultant built her pre-call brief herself in two afternoons, and the course creator’s welcome sequence took a weekend with no developer. Across the 50 cases I keep, 19 builds needed no developer.
How do I know an agentic workflow is working?
Pick one business number before you build, record it for the period before, and keep a log of what the agent decided and what you overrode. Then compare, and write down anything else that changed in the same period.
