From 8–10 working days to 2–8 hours per estimate

CHR’s estimates rested on the knowledge of a few experienced people. That knowledge was excellent, but the process around it was almost entirely manual: CHR had no automated estimating workflow of any kind. Turning a client’s scope documents into tasks, resources and costs took over 65 hours per estimate, and even a small change to a project meant recalculating and redoing much of the work.

Depending on a few individuals also meant estimates varied with who prepared them, and made CHR’s know-how hard to retain. CHR wanted its own estimating norms captured in a system, so that estimates would be consistent and faster to produce while estimators stayed in control of every figure.

“An estimate meant working through the client's documents and pricing every task, crew and piece of equipment by hand, which could take a lot of time, really. The worst part was that a small change in one of the project estimation facts, could cause so much changes along the way, it could mean that none of the work that was done could be reused, and much of the work was started over. When workflow is being done through the software it is mostly double-checking if everything is 100% right, and even if something changes later, it is easy to make needed changes respectively.”

An estimate becomes the price a client sees, so the answers behind it have to hold up. Bloomteq chose Claude for the reliability and quality of its answers, and built the workflow so that Claude drafts and CHR’s estimators confirm. Everything Claude produces passes through an estimator’s review before it reaches a proposal, and the figures themselves come from CHR’s rules rather than from the model.

Bloomteq matched each model to its job. The questionnaire is where a misread document costs the most, so Claude Sonnet 5.5 drafts it, delivering the quality that step needs without the extra cost and latency of a larger model. The supporting steps run often and need speed more than depth, so Claude Haiku 4.5 handles them: analyzing images, enriching the reactor context, asking clarifying questions, matching items to the responsibility matrix and drafting scope-of-work text.

Production use bears this out: in the past 6 months dozens of estimations were done, and estimators now check and adjust Claude’s drafts instead of working through every document themselves.

“We chose Claude over other models because its answers hold up: it fills in over 60% of each questionnaire straight from the client's documents, which is great considering that even those documents don’t contain 100% of needed information, and CHR's estimators keep >90% of those answers unchanged, and even answers that get changed are minor enhancements caused by the extensive domain knowledge of the estimators. Also, the Claude models with their exceptional context window and performance, get reactor questionnaire ready in about 60-180 seconds. And it is also worth mentioning that Claude models successfully mark lacking information, which is as important as extracting facts that are present.”

An estimator starts a project by uploading the client’s scope documents or pasting the scope text, then sets up a separate estimate for each reactor. Document text is extracted directly, with AWS Textract as an OCR fallback for PDF pages that have little selectable text. Claude Haiku 4.5 analyzes images embedded in the documents and enriches the reactor context, and Claude Sonnet 5.5 drafts a structured questionnaire from the source material. Where information is missing or ambiguous, Claude Haiku 4.5 asks the estimator targeted questions.

Once the estimator has reviewed and edited the questionnaire, a configurable rule engine takes over. CHR’s templates, norms and factors generate the task and resource lists, and its rate tables drive the cost breakdown. The rule engine itself makes no model calls, so every figure comes from CHR’s own rules, and a change to the inputs is recalculated rather than reworked by hand. Claude Haiku 4.5 has one more job at this stage: it matches each equipment and supply item on the resource list against the responsibility matrix in the client’s documents, marking whether CHR, the client, both or a third party provides it. Client-supplied items drop out of CHR’s cost total, and the proposal includes the responsibility matrix, so every party knows what it is responsible for. Estimators review the tasks, resources and costs at explicit checkpoints, then edit the proposal sections and download a CHR-branded proposal PDF with supporting PDF and spreadsheet exports.

Bloomteq built the estimation platform as part of a wider rebuild of CHR’s CRM, covering clients, jobs and reports. Using Claude Code, Bloomteq’s team migrated CHR’s entire CRM from the old system to the new one in 2–3 weeks. The platform runs on AWS and synchronizes client and job data with NetSuite.

“From the beginning, we were not aiming towards a system that guesses the numbers and makes calculations unpredictable. CHR's estimators know this work better than anyone, so their rules and decisions are always final and they start from the rulesets used in their business, and Claude does what rules can't: reading each client's documents, gathering necessary facts from the text which is not formatted in any predefined way, spotting what's missing and working out who is responsible for what.”

Production use of the system has proven that preparing an estimate now takes 2 to 8 hours, down from 8–10 working days of manual work. Estimators now start from Claude’s draft and the rule engine’s tasks, resources and costs instead of a blank page, and a scope change means a recalculation rather than a rework. CHR’s estimating know-how no longer sits with just a few people: it lives in the platform’s templates, norms and rate tables. Each proposal also includes a responsibility matrix, so CHR and its client know who provides what.

“Our estimators used to spend too much time building the estimates by hand, even though the estimates should have been very similar most of the time, and every time the change occurred, it could mean starting over. Now they start from a draft that already follows our rules, and embeds all the knowledge and rules that are applied throughout the process of estimating a project, and the experience of our experts finally goes into reviewing the job instead of rebuilding the numbers. What I highly value about the new way of doing estimations is the fact that we finally can have branded and unified proposals, but still tailored to the customer’s needs, without putting any effort into it, relying fully on the workflow that was made.”