Spend cube delivery for procurement consultancies

Quick answer: For a procurement consultancy that builds spend cubes by hand for its clients, the constraint is not skill, it is analyst hours. An AI classification pipeline replaces the week of manual categorisation with a taxonomy that cannot invent categories, so a consultant can review flagged lines instead of coding all of them, and price the work as a fixed fee up to a threshold plus token cost beyond it.

On this page: Why manual cube building limits a consultancy · Taxonomy guardrails against hallucinated categories · A fixed fee or token cost model · What this changes for client delivery · FAQ

A consultancy that builds spend cubes for clients sells judgement, but delivers a lot of it by hand. A senior consultant sits with a spend export, works line by line against a taxonomy, and writes down a category for each row. That work is billable, and it is also the reason a consultancy can only run so many of these engagements in a quarter.

The constraint is not client demand. It is that classifying spend at the line level was, until recently, something only a trained person could do reliably. That is no longer entirely true, and it changes two things at once: how the work gets delivered, and how it gets priced.

Why manual cube building limits a consultancy

For a consultancy, the spend cube is one deliverable inside a larger engagement, sourcing strategy, supplier negotiation, category management. It is also the deliverable that eats the most hours for the least judgement. A senior consultant does not need ten years of category management experience to decide that a line reading "filters, various" in an HVAC client's export belongs in air filtration rather than water treatment. They need the time to sit with the export and work through several thousand lines like it, one at a time.

That time does not scale with the number of clients a consultancy can take on. Each new client means a new taxonomy pass, a new set of edge cases, a new week or two of an experienced person's calendar spent on classification rather than on the negotiation or the category strategy the client is actually paying for.

The Copilot shortcut does not fix this, it just moves where the hours go. A consultant who points a general assistant at a spend export gets an answer quickly and then spends longer checking it, because the tool invents categories that do not exist in the client's taxonomy and classifies the same line differently if asked twice. The bottleneck does not disappear. It becomes a review problem instead of a build problem.

Taxonomy guardrails against hallucinated categories

A consultancy's exposure here is different from a company classifying its own data. When an in-house team gets a category wrong, they fix it quietly. When a consultancy hands a client a spend cube with an invented UNSPSC code sitting in a slide, the consultancy's name is on that slide, not the tool's.

That is the reason a taxonomy has to be constrained before it can be trusted for delivery work. A constrained taxonomy means the classification step can only select from the codes that exist in the standard, or in the client's own category tree if one has already been agreed. It cannot generate a plausible-sounding code that is not on the list, because the list is not a suggestion. It is the entire space the model is allowed to search.

The second guardrail is confidence scoring. A description that is only a part number, a supplier appearing for the first time, an item that could sit in two families with no way to tell from the text alone, gets flagged rather than guessed at. Layered guardrails of this kind, grounding the model in a fixed reference list, scoring confidence, and routing uncertain cases to review, have been shown in recent evaluations to cut hallucination rates by roughly 70 to 90 percent compared with an unguarded setup. That gap is the difference between a consultant spot checking a sample and a consultant re checking everything, which is the same manual week the pipeline was meant to remove.

For a consultancy, guardrails are not a technical detail. They are the reason the classification step can run unattended between the moments a person actually needs to look at it.

A fixed fee or token cost model

The economics of a spend cube engagement have followed the economics of the work: a project fee scoped to the analyst hours it takes to code the file. That model made sense when classification was manual and the price had to cover a person's calendar for two to four weeks.

A pipeline changes what the fee is actually paying for. The marginal cost of classifying the next thousand lines is close to zero once the taxonomy and the rules are set, so a straight hourly or day rate structure stops reflecting the actual cost of delivery. It also stops rewarding the consultancy for building the reusable part, the rules, the supplier resolution logic, the taxonomy decisions, because those get thrown away and rebuilt for the next client under the old model.

A cleaner shape is a fixed fee up to a line item threshold, with a per token or per line cost above it. A client with 40,000 transaction lines is a predictable, fixed price engagement. A client with 400,000 lines, or one where classification keeps running month over month as new invoices arrive, moves into usage based pricing that scales with volume instead of a renegotiated day count.

This is not a live product feature, it is a pricing shape a consultancy can build its own commercial model around, using a pipeline underneath it. It works because it separates two costs that a day rate blends together: the fixed cost of setting up the taxonomy and rules once, and the variable cost of running classification against however much data a given client has.

What this changes for client delivery

When classification is a pipeline rather than a person's week, what a consultancy sells shifts too. The spend cube stops being the billable centrepiece of the engagement and becomes an input to it, produced faster so the consultant's time goes into the category strategy and the supplier negotiation the client actually hired them for.

It also changes what happens after the engagement ends. A manually built cube is accurate on delivery day and starts decaying the moment new invoices arrive, because nobody at the client holds the rules the consultant used. A pipeline can keep classifying new lines against the same taxonomy after the project closes, which gives a consultancy a reason to stay attached to the account instead of handing over a static spreadsheet and moving to the next tender.

Pearstop runs that classification layer for consultancies who want to keep their delivery model and stop billing analyst hours for the coding step, freeing the consultant to sell the strategy work a spreadsheet was never going to do.

For a consultancy weighing this up, the decision is not whether AI replaces the judgement work. It does not. It is whether the classification step, the part that was never really judgement to begin with, keeps consuming a senior person's week on every engagement, or becomes infrastructure the consultancy runs underneath the work it was actually hired to do.

Frequently asked questions

Can a procurement consultancy offer spend cube building as a service using AI?

Yes. A consultancy can run spend classification through an AI pipeline instead of manual coding, then apply its own judgement to the strategy and negotiation work around it. The classification step becomes infrastructure the consultancy runs under its own name, with a constrained taxonomy and human review on the lines the system is unsure about, rather than a purely manual build.

How do you stop an AI classification pipeline from hallucinating categories?

By constraining what the model is allowed to output. A pipeline that can only select from the fixed codes in a taxonomy, rather than generate a plausible sounding one, cannot invent a category that does not exist. Paired with confidence scoring that flags uncertain lines for human review, guardrails like this have been shown to cut hallucination rates by roughly 70 to 90 percent compared with an unguarded setup.

What is a fixed fee plus token cost pricing model for spend cube work?

It is a pricing structure where an engagement up to an agreed line item threshold is billed at one fixed fee, and any volume above that threshold is billed per token or per line classified. It replaces a day rate model that charges for analyst hours with a structure that reflects the actual cost of running a pipeline: a fixed setup cost plus a variable cost tied to data volume.

How does a consultancy trust AI classified spend enough to present it to a client?

Trust comes from the guardrails, not the model itself. A constrained taxonomy stops the system inventing codes, confidence scoring separates lines it is sure about from lines it is not, and uncertain lines go to a person before anything reaches a client deliverable. A consultancy is checking the flagged exceptions, not re verifying every line, which is what makes the output defensible under its own name.

How does Pearstop help procurement consultancies build spend cubes faster?

Pearstop runs the classification layer that sits under a consultancy's delivery work: a constrained taxonomy so codes cannot be invented, supplier resolution, and confidence scoring that routes uncertain lines to review. Consultancies use it to produce the classified spend cube in less time than a manual build, and spend their own hours on the category strategy and negotiation the client actually engaged them for.

Free resources

• Pearstop case studies
• Procurement Opportunity Mapper
• Taxonomy generator, coming soon

Stephanie Wiechers

CEO & Co-founder, Pearstop

Stephanie leads Pearstop's go-to-market and strategic direction. She works directly with procurement and FM leaders across Europe to understand how data quality affects margins, contracts, and AI readiness.

LinkedIn →