Most of what you want is automation, not an agent.

Knowing which one a job needs is the difference between something that runs quietly for a year, and something that costs more every month and cannot explain itself.

Send me one problem Free answer, even if the answer is do not automate it
Bahgat Ahmed
7 years shipping to production
LEAD IN→ ENRICH→ ROUTE→ REPLY→ BOOKED +WATCH

Everyone builds the first five. The last one is the one that tells you when the first five stopped.

what actually happens

A workflow that dies does not send an error.

It just stops. The sheet keeps showing last week, the follow ups quietly stop going out, and the first person to notice is usually a customer. Here is what that looks like on the canvas.

 lead-follow-up · last successful run 3 days ago
New lead webhook Enrich lookup Send reply token expired nobody was told Book never ran Log to CRM never ran Three days. Zero alerts.
the distinction that decides your bill

Automation and agents are not the same product.

They get sold as one thing. They fail in completely different ways, and one of them charges you every time it runs.

Automation

You decide the path

Written down in advance. Same input, same route, every time.

step 1 step 2 step 3
  • Predictable. You can read the whole thing on one screen
  • Cheap to run. No model call for a decision a rule already makes
  • Auditable. When it goes wrong, it goes wrong in one visible place
  • Breaks when input arrives in a shape nobody planned for
Agent

The model decides the path

It picks which tools to use, and in what order, while it runs.

model decides search the CRM read the PDF ask a human
  • Handles the unplanned. Input with no fixed shape, output needing judgement
  • Costs predictability. Two identical requests can take different routes
  • Can be confidently wrong and report success, with no error anywhere
  • Priced per run, so the bill grows with use instead of sitting still

The choice is made per step, not per project. Most systems are automation with one or two model steps inside them, and that is usually the cheapest thing that actually works.

how I decide

Three questions, in this order.

1

Fixed shape in, fixed shape out? Write the rule. Clocks, comparisons, moving data between systems. Automation already does this, and it does it better because it never has an opinion.

2

No fixed shape, or a judgement needed? Use a model for that one step, and add a check. A model buys you exactly two things a rule cannot do. Everything else is a rule wearing a costume.

3

Can the path itself not be written in advance? Only then does an agent earn its place. If you can draw the flowchart, you do not need something to invent it at runtime.

the part that is usually missing

So I build the system, and the thing that watches it.

A heartbeat on every workflow, a freshness stamp on every row, and one message when something stops moving. It costs almost nothing to add and it is the difference between finding out on Tuesday and finding out from an angry customer on Friday.

You get the message. Not your customer.

✓

CRM sync

Ran 06:00. 42 rows written.

!

Scorecard has not updated

Newest row is 41 hours old. Expected daily.

✕

Quote chaser stopped

No run in 3 days. Last error: token expired.

what this usually looks like

The same few systems, in different clothes.

Different industries, one shape underneath. Something happened, and a person has to remember to act on it.

01

Speed to lead

Enquiry lands, reply goes out in seconds, qualified, booked.

usually no watcher
02

Quote follow up

The chaser that does not forget on day three, or day ten.

usually no watcher
03

CRM and sheet sync

Your scorecard updates itself, keyed so a rerun does not duplicate.

usually no watcher
04

Client reporting

Assembled, formatted and sent on a schedule.

usually no watcher
05

Support triage

First reply handled, the hard ones escalated to a person.

usually no watcher
06

Document intake

PDFs and emails into records, with approval before anything commits.

usually no watcher
who you would be working with

Bahgat Ahmed

Bahgat Ahmed

Seven years across enterprise and consumer facing AI, most of it for government and regulated clients where a quiet failure was a real problem rather than an inconvenience. I rebuild stuck AI products and ship them to production. Not demos, not prototypes.

Working with smaller businesses taught me the more useful half. In a small company the process usually lives in one person's head, so the question is almost never which tool. It is where does this break, and who finds out. That is why I look at the whole flow before building any part of it.

n8nMakePython Claude & OpenAIRAGAPIs 1B Lab

I write about all of this, at some length.

Send me the one thing that keeps getting forgotten.

Tell me the job someone on your team has to remember to do. I will tell you whether it is a rule, a model, or an agent. Sometimes the answer is that it should not be automated at all, and that answer is free.

bahgat@goboldai.com

No call needed to start. One paragraph is enough.