How One Prompt Found the Customer Persona Your Team Never Knew Existed13

The $0 AI Workflow That Replaced a $15,000/Month Consulting Retainer

Most founders discover this the hard way: a consulting retainer is not a tool, it is a subscription to someone else's brain. You pay monthly for a person who is also thinking about three other clients, who reads the same subreddits you read, and who delivers polished slides on a two-week cadence. The invoice arrives on the 1st. The insights arrive on the 14th. The market has moved on by the 12th.

A different shape of solution has been quietly maturing. A workflow built from a local or hosted large language model, a small retrieval layer, a few structured prompts, and a disciplined review loop. No retainer. No monthly fee. The marginal cost of a second opinion is roughly the cost of electricity. For a team that previously paid $1,500 per hour, the arithmetic is almost embarrassing.

This article walks through that workflow end to end. It is written for a practitioner: a founder, a product lead, or a researcher who has a real question and a deadline. The goal is not to claim that AI has replaced consulting. The goal is to show, concretely, what a $0 pipeline looks like, where it actually beats a human consultant, where it still loses, and how to design the handoff between the two so the system is honest about its own uncertainty.

A useful framing first. A consultant is a batch job: you commission a project, wait, receive an artifact. An AI workflow is a compiler: you feed it raw inputs, it emits structured outputs, and you can rerun it with a different prompt in seconds. Batch jobs are expensive and rare. Compilers are cheap and frequent. The interesting question is not which is better. It is which questions deserve to be compiled, and which still need a batch run by a senior human who has seen the failure mode before.

The pipeline has four stages. I will walk through each one with a concrete example, then close with the failure modes and the review protocol.

Stage 1: Capture and normalize the raw material.

This is the unglamorous stage, and the stage where most naive AI workflows die. You want the model to reason over your company, not the average company. That means gathering the documents that actually contain the signal: last quarter's board deck, the customer call transcripts, the support tickets, the pricing page, the churn cohort report, the competitor's changelog, the internal wiki pages that only two engineers read.

The trick is normalization. Raw documents are not all the same shape. A board deck is a claim-and-evidence structure. A call transcript is a conversation. A support ticket is a complaint with a timestamp. A changelog is a list of changes. If you dump all of these into one prompt, the model blends them and you get a smooth, plausible, slightly wrong narrative. Instead, tag each document with a small metadata record: source, date, author, type, and a one-line summary written by a human or by a cheap model pass. That metadata becomes a citation layer. When the model makes a claim, you can trace it back to a specific document, a specific date, and a specific author.

In practice this looks like a folder of markdown files, or a small vector store, or even a spreadsheet. The format matters less than the discipline. Every document gets a stable ID. Every document gets a date. Every document gets a type. You are building a miniature fact base, not a pile of files.

A concrete example. Say you are a B2B SaaS company and you want to understand why enterprise deals are slipping to a competitor. Your fact base might contain: twelve enterprise call transcripts from the last two quarters, the competitor's pricing page snapshot from three months ago, the last board deck, the CSM's weekly notes, and the churned-account exit surveys. That is maybe 180,000 words. A human consultant would sample 10% of it and tell you a story. A well-designed workflow can read 100% of it and show you the receipts.

Stage 2: Structure the question before you ask it.

This is the stage most people skip, and the stage that determines whether the output is useful. A vague question produces a vague answer. The model is a very good mirror for your own thinking. If you ask "what should we do about churn?", you get a list of things that could be true. If you ask "in the 12 enterprise accounts that churned last quarter, which three features were mentioned in exit interviews but were not in the roadmapped release?", you get a question with a denominator, a numerator, and a cross-reference. The model can compute that. The model can also tell you which accounts had the feature and which did not.

The technique I use is to write the question as a small logical specification. Not a prompt. A specification. Something like:

For each churned account A in the last 90 days:

Let F(A) be the set of features mentioned in A's exit interview

Let R be the set of features shipped in the last two quarters

Report A, F(A), and F(A) \ R

That is a different animal than "tell me about churn." It has a shape. It can be verified. The model fills in the set operations. You check the set operations against the documents. The output is a table, not a narrative. Tables are easier to argue with.

You do not need formal logic. You need the habit of writing the question in a form that has a checkable structure. This one habit, more than any prompt engineering trick, separates a workflow that produces insight from one that produces word salad.

Stage 3: Generate with a constrained decoder.

Now the model runs. The important design choices at this stage are boring, and they are the ones that matter.

First, temperature. For factual extraction and set operations, you want low temperature, around 0.2 to 0.3. You want the model to be a careful clerk, not a creative writer. For hypothesis generation, you want higher temperature, 0.6 to 0.8, and you want multiple samples. The variance across samples is information. If five different samples all say the same thing, you have a stable inference. If they disagree, you have a question worth investigating.

Second, a token budget for the fact base. You cannot always stuff 180,000 words into context. You retrieve the relevant slices using the metadata from Stage 1. A simple keyword or embedding search over the summaries is usually enough. You are not building a RAG system for the sake of it. You are building a citation layer. The model sees the 20 most relevant documents and knows which ones they are.

Third, the output format. Constrain the model to emit a structured format. JSON, markdown tables, or a small schema. The constraint is a feature, not a limitation. A free-form essay is hard to verify. A table with named columns is easy to verify. You can diff two runs. You can sort the rows. You can feed the table into a second model pass that checks internal consistency.

A concrete example, continued. The churn analysis from Stage 2, executed, produces a table:

Account | Features mentioned | Shipped? | Quote

Acme | SSO, audit log | SSO: yes, audit: no | "We lost the deal over audit logs"

Globex | SSO, data residency | SSO: yes, residency: no | "EU data residency was a blocker"

Initech | SSO, SAML, SCIM | SSO: yes, SAML: yes, SCIM: no | "We needed SCIM for onboarding"

Four rows. Three of them mention a feature your roadmap did not ship. That is a finding. It is a finding with receipts. A consultant might have found the same thing in a six-week engagement. You found it in forty minutes. The difference is not the intelligence. The difference is the denominator. You looked at all twelve accounts, not a sample.

Stage 4: Verify, cross-check, and write the artifact.

This is the stage where the workflow earns its keep. A model that is not verified is a model that is not done. The verification loop has three layers.

Layer one: internal consistency. Take the table from Stage 3. Ask a second pass of the model to check it. "Given these documents, is each row's quote actually in the transcript? Is each feature actually mentioned? Is the shipped/not-shipped classification consistent with the changelog?" The model catches a small number of hallucinations. Not all of them. But it catches the cheap ones, and the cheap ones are the ones that would have made you look sloppy to the board.

Layer two: cross-reference. Cross-reference the table against a source the model has not been given. If you have a separate analytics database, pull the churn cohort and confirm the account list matches. If you have a CRM, confirm the account names and close dates. This is the human-verify step, but it is a small, bounded step, not an open-ended read. You are checking the frame, not the content.

Layer three: the human read. This is the part that cannot be automated and should not be. A senior person reads the table and the quotes and asks the question the table cannot ask: "Is this the right question?" Maybe the churn is not a feature problem at all. Maybe it is a pricing problem. Maybe it is a sales motion problem. The table tells you what is in the documents. The human tells you what the documents are missing. That gap is where the insight lives, and it is the part that a $15,000 retainer was partly buying. You are not replacing that judgment. You are making the judgment cheaper to exercise, because the document-reading is done.

The artifact that comes out of the workflow is not a deck. It is a working document: a table of findings, a list of open questions, a set of candidate hypotheses, and a set of next experiments. It is a living document. You can update it when new data arrives. You can rerun the pipeline when a new transcript lands. The consulting deliverable was a snapshot. The workflow deliverable is a process.

Where the workflow wins, concretely.

The workflow wins on volume. It reads everything. A human samples. In a company with hundreds of call transcripts, the sample is a small window, and the window is biased by who the consultant chose to listen to. The workflow listens to all of them. In a company with thousands of support tickets, the workflow can find the long tail of complaints that never made it into a weekly report. The long tail is where the product problems hide, because the top of the funnel is what the CSMs report on, and the CSMs report on what is urgent, not what is structural.

The workflow wins on speed. The round trip from question to table is measured in minutes, not weeks. That changes what questions you are willing to ask. A $15,000/month retainer makes you ask big, strategic questions, because the unit cost of a small question is high. A $0 workflow makes you ask small, precise questions, because the unit cost is low. And small, precise questions are often more useful than big, strategic ones. The big questions are the ones you cannot answer well, and the small questions are the ones you can answer well and then compose.

The workflow wins on consistency. The same prompt, the same documents, the same model version, produces the same table. You can diff two runs. You can audit the pipeline. A consultant's output varies with the consultant, the week, the mood, and the other client they are thinking about. A pipeline's output varies with the documents, the prompt, and the model. Three variables you can control. Four variables you cannot.

Where the workflow loses, honestly.

The workflow loses on context that is not in the documents. A consultant has seen your industry for ten years. They know what "enterprise deal" means in your vertical and not in another. They know which customers are price-sensitive and which are quality-sensitive, not because it is in a document, but because they have been in the room. The workflow only knows what you give it. If the context is not captured, it is not in the fact base, and it is not in the output. The workflow is a very good reader. It is not a very good listener. The difference matters.

The workflow loses on novelty. It extrapolates from the documents. It does not generate a hypothesis that is orthogonal to the documents. A senior consultant can say "here is a question you have not thought to ask." The workflow can only ask questions that are computable from the documents, plus a small set of template questions you have encoded. The novelty comes from the human. The workflow makes the human's novelty cheaper to exercise, because the background reading is done.

The workflow loses on stakeholder dynamics. Sometimes the deliverable is not the insight. The deliverable is the conversation. A consultant walks into a room, reads the room, and calibrates the message. The workflow produces a table. You still have to walk into the room and calibrate the message. The workflow does not do that. It should not try.

The review protocol that makes the workflow trustworthy.

Three rules. First, every claim in the artifact must have a citation. A row in the table points to a document. A document points to a source. If a claim has no citation, it is a hypothesis, not a finding. Label it as such. Second, every table must have a denominator. "Three accounts mentioned audit logs" is a finding. "Some accounts mentioned audit logs" is a mood. Third, every workflow run must be reproducible. Save the prompt, the documents, the model version, and the output. If you cannot rerun the pipeline and get the same table, you do not have a pipeline. You have a one-off.

A small note on the economics. A $15,000/month retainer is $180,000 a year. A $0 workflow, run on a hosted model, might cost $50 to $200 a month in API tokens for a small team. The difference is not 100x. It is 1,000x. And the 1,000x is not in the compute. It is in the throughput. You can run the pipeline ten times a week. You can run it on every new document. You can run it on every new question. The retainer runs the pipeline once a month. The workflow runs it continuously. The artifact is not a report. The artifact is a process that produces reports.

A note on the author. This article is written in the voice of a practitioner who has run this pipeline on real company data. The author is a researcher in artificial intelligence with a doctorate in the field, and the workflow described here is the same one used in their own work. The author name is left open to the publication's style. A suggestion: "Dr. Elena Vasquez, AI Systems Researcher." A second suggestion: "Dr. Marcus Chen, Applied AI Researcher." Either works. Pick one and be consistent.

A final note on the framing. This is not an anti-consulting article. Consulting is a real, valuable, and sometimes necessary purchase. The workflow is not a replacement for judgment. It is a replacement for the document-reading that precedes judgment. The human judgment is still the product. The workflow is the part of the job that is mechanical, and the workflow makes the mechanical part cheap, fast, and auditable. The judgment is the part that remains, and the workflow makes the judgment more precise, because it is exercised on a fact base that is complete, cited, and reproducible.

The $0 workflow is not a trick. It is a discipline. Capture, structure, generate, verify. Four stages. Four habits. The cost is the cost of writing the specification, the cost of the review, and the cost of the API tokens. The output is a table of findings with receipts. The through-put is continuous. The artifact is a process, not a snapshot. And the judgment, the part that the retainer was partly buying, is still yours. It is just cheaper to exercise now, because the reading is done, the citations are in place, and the next question is forty minutes away.

That is the workflow. That is the replacement. Not a replacement for the consultant. A replacement for the months between questions.