Experiment · EXP-001

Can an LLM own the entire quoting workflow?

End-to-end generation versus interpretation plus deterministic pricing in QuoteSnap AI.

The tempting architecture for QuoteSnap was a single model that inspects the job and emits a price. I treated that as an experiment and abandoned it as the owner of contractor economics. The model remained useful for interpretation. Pricing stayed in software. This experiment is qualitative and production-derived, not a published accuracy trial.

Hypothesis: A multimodal model can inspect job details and images and directly produce a trustworthy quote.

Research question

Can a multimodal language model safely own an end-to-end quoting workflow, from interpreting the job to deciding the final amount a contractor should charge?

Why I tested it

The simplest AI-first architecture for QuoteSnap was also the most tempting: send the customer description, property context, uploaded images, and contractor preferences to a multimodal model and ask it to return a finished quote. It reduces orchestration and makes the model appear to own the whole workflow.

Setup

I evaluated that model-centred approach against a layered architecture. In the layered version, the model interprets ambiguous inputs such as customer language and property images, then emits structured job signals. Those signals are validated before a deterministic pricing engine applies contractor-defined rules such as minimum charges, job-type pricing, add-ons, percentages, and location-specific adjustments.

Observation

The model was useful where the input was uncertain. It could classify work, interpret a description, and turn visual information into structured signals. The problem appeared when the same probabilistic component was also allowed to own contractor economics. A final price is not merely an interpretation. It is the result of business rules that should remain stable, inspectable, and editable.

Result

I abandoned the model-owned pricing architecture. QuoteSnap retained AI for interpretation and kept pricing in deterministic software. This created a cleaner failure boundary: if a quote is wrong, I can ask whether the image interpretation failed, whether validation failed, whether the contractor configuration is wrong, or whether the pricing engine contains a bug.

What survived

The useful pattern became: AI interprets, software decides, AI may communicate. The model remains valuable, but it no longer owns decisions that should be repeatable under the same business configuration.

I trust AI to look at a dirty driveway. I do not trust it to decide what that driveway should cost.

Expected: A single model call would inspect the job and emit a usable price.
Actual: The model could interpret the driveway and still be the wrong owner of contractor economics.
Cause: Capability was treated as permission. Pricing needs stable rules, not a new sample from a distribution.

Limitations

This is a production-derived qualitative experiment, not a published benchmark comparing quote accuracy across a labelled evaluation set. The result supports an architecture decision for this class of product. It should not be read as a universal claim that models can never participate in pricing.

Abubakar, Imam. "Can an LLM own the entire quoting workflow?." Imam Abubakar AI Research, 2026. https://imamabubakar.com/research/exp-001-llm-owned-quoting

This archive contains independent research, experiments, architecture studies, and technical observations. Some findings are exploratory and should not be interpreted as peer-reviewed scientific conclusions unless explicitly stated. Architecture studies of third-party products are based on publicly observable behaviour, published technical information, and engineering analysis. They are not claims of access to private source code or infrastructure.