Skip to content
•
10 min read

AI consulting start to finish: Nick Saraev's six steps

Talk

How I'd Make $12,373/Mo as an AI Consultant in 2027 (Course) by Nick Saraev, his YouTube channel, Oct 29, 2026 (28 min)

Watched

These are my notes on what the speaker said. I've checked the names, numbers and quotes against other sources.

Key ideas

  • A business moves at the speed of its slowest step, so the first job is finding that step, and work anywhere else changes nothing.
  • For most small businesses the slow step is in growth: getting seen, getting interest, closing, or earning more from existing customers.
  • Break the goal of fixing the bottleneck into smaller and smaller parts until each one is an action an agent or a person can take.
  • Hand over a system the client runs without you, in their own account, so you aren't the next bottleneck.
  • An AI system gives a spread of outputs, so run it many times and score the results before you trust it or change it.
  • A weekly call where the best result of the week gets turned into a shared skill moves what works from the strongest users to everyone else.

Nick Saraev’s method for AI consulting has six steps, and only one of them is building. Find the one bottleneck in the business, break the goal of fixing it into small concrete actions, build the system with a coding agent, deploy it so it runs without you, test it with evals, and set up a weekly call where the team shares what works. Then the bottleneck moves and you start again. These are my notes on his 28-minute video, with the questions he asks on a first call, a worked example for each step and the three mistakes he warns about. Revenue figures and client stories are as he tells them; the survey and product facts are checked against their sources.

The six steps

StepWhat you doWhat you have at the end
1. Find the bottleneckAsk for a number at each growth stage on the first callThe one stage worth working on
2. Goal factoringBreak the goal of fixing it into a tree of actionsA list of small, concrete tasks
3. BuildGive the tasks and the client’s accounts to a coding agentA system that worked once
4. DeployTurn it into a skill and a scheduled routine in the client’s accountA system that runs without you
5. EvalsRun it many times, score the outputs, rerun after changesA score you can improve against
6. Team syncA weekly call that turns the best result into a shared skillA team that keeps it current

He puts building at about a sixth of the job and says the rest takes common sense more than technical skill. His case for the demand is that most businesses have barely started. The Census Bureau’s Business Trends and Outlook Survey had 17% to 20% of US businesses using AI in any business function between December 2025 and May 2026, and that counts any use in the previous two weeks.

Finding the bottleneck

The idea comes from Eliyahu Goldratt’s 1984 novel The Goal. A system only moves as fast as its slowest step. Speed that step up and the whole business produces more. Speed up any other step and nothing changes, because the slow one still sets the pace. His example is a business with plenty of leads that takes two or three days to reply to them. Better advertising wouldn’t help, because the constraint is the reply.

He looks for the bottleneck in four stages he calls RACE:

StageWhat it measuresA sign the bottleneck is here
ReachHow many people see the businessToo few people see it for anything later to happen
AcquireHow many of them show interest: a booked meeting, a form filled inLots of attention, few meetings
ConvertHow many interested people payMeetings happen and few close
ExpandHow much each customer pays over timeCustomers leave early or never buy more

He stays on the growth side on purpose. Most small businesses, he says, don’t have trouble delivering. They don’t have enough volume for delivery to be the problem.

On the first call he goes through the four stages one at a time and asks for a number at each. The stage with the worst rate, measured against what’s normal for that kind of business, is almost always the bottleneck. In his example a marketing agency gets 1,000 views a day, books 10 meetings and closes 1. A 1% rate from views to meetings is normal for an agency that size, as he tells it, and closing 1 in 10 is well under the 25% to 30% he treats as typical. So every conversation after that is about the sales process: training, follow-up, scripts. Nothing else gets his time.

The reverse case is the consultant who picked a service because it sounded impressive. If it isn’t the client’s bottleneck, nobody buys it.

Goal factoring

Once you know the bottleneck, state fixing it as a goal and ask how you’d achieve it. Then ask the same of each answer, and keep going down until you reach actions someone can take today. He says it takes about 15 minutes of hard thinking and a diagram.

His example starts from a reach problem:

  1. Goal: more people see the ads.
  2. Ways to get there: raise the budget, add a second ad platform, give the current platform more ad creative, write more ad copy.
  3. Under more ad creative: find the ads that already work, write new hooks and headlines for them, remake the video ads in new formats.
  4. Under find the ads that work: pull the account’s results and rank them.

What comes out is a long list of small tasks, and many of them suit an agent. He gives the whole diagram to a model and asks for a plan to do as many as it can.

He credits the name to the Center for Applied Rationality. Their goal factoring runs the other way: it starts from something you already do, lists the goals it serves, and looks for better ways to meet them. What he describes is a top-down goal tree.

Building the system

The agent does most of the building. In the ad example he connects it to the client’s ad account through a connector. Meta opened one for AI assistants in beta in late April 2026, and it can read reports and create ads. His prompt has this shape:

  1. Pull a report on every active ad from the last 90 days and group the ads by type.
  2. Rank them by cost per lead.
  3. For each top ad, write five new hooks, five headlines, five versions of the main text and a few variations of the offer.
  4. Look through Meta’s public ad library for competitor ads worth learning from.
  5. Create the new ads in the account as paused, so nothing spends money.
  6. Summarize what was made and why.

Then he checks the output himself before anything goes live, keeps what works and turns off what doesn’t.

He’s plain that some of this is easy. Part of a consultant’s value, as he sees it, is being outside the business. The owner is too deep in the daily work to step back and ask what the real problem is.

Deploying it

A one-off run isn’t the product. The client should be able to run the system without calling you, for two reasons: the system is something they own, and you don’t want to become their bottleneck.

How he deploys depends on whether a model is part of the flow:

The flowWhere it runs
No AI in it, only scriptsA cloud host for scripts. He names Modal and Trigger.dev, and has the agent do the upload.
AI in itA routine: the run that worked becomes a skill, gets its connectors and runs in the cloud on a schedule or a trigger.

Routines are a Claude Code feature, and Anthropic’s docs now also call them cloud scheduled tasks. He says a consultant can build 20 or 30 for one client, each doing work you can point to.

He sets all of it up inside the client’s own team or enterprise account, as an administrator or on a shared login for the setup period, because the staff will need to use and check the automations. In the sales conversation he says so directly: you get all the systems, and they should run with or without me.

Evals

A model doesn’t give one output. It gives a spread, and one good run says little about the next five. A good automation has a narrow spread, and the only way to see the spread is to run it repeatedly.

His version is simple. Run the system ten times, rate each output out of ten, and keep the average. After any change, run the test again and compare. For larger clients he reruns every skill and procedure whenever a new model comes out, and improves in small steps. The extra usage costs little, he says, next to the cost of not knowing.

The weekly team sync

In any team some people use AI much better than others, and who is ahead varies by task. He sets up one weekly call so that what the best people do reaches everyone else, and then stops attending.

PartHow he runs it
SizePods of 5 to 12. Above about 50 staff, one call per department.
Who presentsWhoever got the best result on last week’s tasks walks through exactly what they did.
What gets keptTheir chat logs go to an agent, which works out the procedure they followed and writes or updates a skill.
How it spreadsThe skill is pushed to everyone through the team plan.
How oftenWeekly. He finds more often than that isn’t worth it.

Three mistakes

MistakeWhat he does instead
Automating a process that isn’t the bottleneck, or is brokenSpend nearly all the effort on the four growth stages, at the bottleneck
Inventing a new process nobody ownsTake one that already works at a fraction of its potential and improve it
Working where nothing is measuredAgree at the start that everything gets measured, so the result can be shown

The third one is about getting paid. Results take a while to show up as profit, and he says he’s had clients ask after two months what he’d done. A baseline number from before the work is the answer.

Fixing one bottleneck usually makes another step the slowest. Goldratt’s last rule is not to let habit keep you working on the old one, so he reviews where the bottleneck is on a regular schedule, weekly in his own practice.

How to try it

  • Ask for four numbers: how many people see the business, how many show interest, how many pay, and how long they stay or how much they spend.
  • Compare each rate with what’s normal for that kind of business. Work on the worst one and leave the rest.
  • Spend 15 minutes drawing the goal tree before opening any tool.
  • Have the agent create things paused or as drafts, and review before anything goes live.
  • Run it ten times and write down the average score before changing anything.
  • Write down the starting number, so there’s something to compare against later.

Key terms

Bottleneck
The one step that limits how much the whole business can produce. Also called the constraint.
RACE
His four growth stages: reach (people who see the business), acquire (people who show interest), convert (people who pay), expand (what each customer pays over time).
Lifetime value
The total a customer pays over the whole time they stay.
Goal factoring
His name for breaking a goal into a tree: ask how to achieve it, then how to achieve each answer, until you reach concrete actions.
Connector
A link that gives an AI assistant access to an outside account, such as an ad account, a customer database or a shared drive.
Skill
A written procedure an agent follows, so a task that worked once is done the same way next time.
Routine
A saved agent task (a prompt, its connectors and a schedule or trigger) that runs in the cloud without anyone starting it.
Eval
A repeatable test that scores a system's output, so you can tell whether a change made it better or worse.
Hill climbing
Improving a system in small steps, keeping each change only if the score goes up.

Quotes from the talk

Only one of the six steps... involves the actual building.

Nick Saraev, quoted in the talk

You want to pick something that actually hurts.

Nick Saraev, quoted in the talk

We don't really want to be the bottleneck ourselves.

Nick Saraev, quoted in the talk

Connects to

Other entries that share a theme with this one.

More research

All research

Send feedback

This goes to my inbox through Web3Forms, with this page's address attached. The form sets no cookies.