Automation & AI 8 min read · published on 30 October 2026 · updated on 30 October 2026

AI-assisted estimates: what should you check before sending?

Wail Saidani Wail Saidani Founder of DevDream
Summarize this article with: Claude ChatGPT Perplexity Le Chat Grok
Illustration of a large proofreading desk and a brass stamp held back by a lever in a warm miniature print shop.
A visual metaphor: prepare the document, then decide whether to send it.

An AI-assisted estimate should remain a draft until the work, quantities, prices and terms have been checked. Before sending, approve the exact document and its recipient. If either changes afterwards, approval must be given again: a review does not cover later versions.

A well-presented document can hide a missing material or item. I would start with this question: where does the information in each line come from, and who can confirm it? The attractive layout can wait its turn.

AI-assisted estimates: what can you delegate before pricing the work?

Let the tool prepare descriptions and reuse your verified data. Keep the decisions about the work, quantities and proposed price in your hands. Generative artificial intelligence, which produces text in response to a request, can write a convincing explanation using false information. It must therefore flag unknowns rather than turn them into confirmed work items.

  • What the tool can prepare: organize your notes, reuse your item descriptions and flag missing information.
  • What you need to confirm: site measurements, constraints, materials and applicable prices.
  • What your software should calculate: totals, discounts and value added tax (VAT), with the settings checked.

Missing information should remain visible as a question. In an illustrative window replacement example, removing the old frame does not automatically become included work. It still needs to be confirmed against the work actually planned.

The French Directorate General for Enterprise (DGE) offers this reminder in Violette Jannot's analysis of 7 October 2026 (translated from French):

"the reliability of results, transparency of tools and continued human approval of important decisions remain essential." Read the analysis.

If inquiries are still arriving in scattered places, start by organizing the information before pricing the work. A draft cannot make up for a missing site visit.


How should you check each line before approving?

Link each line to a specific source: a site visit record, supplier reference or current internal price list. Then check what that source actually confirms. An old estimate may provide a useful description without supporting today's price. The following checklist helps distinguish information you have from decisions you still need to make.

Illustration of an estimate connected by copper pipes to three drawers marked with a ruler, samples and a clock in a miniature workshop.
A visual metaphor: verifiable information behind every line of the estimate.

Checklist for reviewing estimate line items

ItemSource to findCheckWhat to do if unsure
Work itemSite visit notes and client request.Included, excluded or offered as an option?Clarify before pricing.
Quantity and unitVerified measurements or plan.Correct dimension, area and unit?Check the measurements again.
Material or item suppliedSupplier reference.Correct product and required accessories?Confirm the reference.
Price and discountCurrent internal price list.Price approved for this job?Have the amount confirmed.
Totals and VATSoftware and applicable rules.Calculations and settings checked?Correct before issuing.
TermsYour checked template.Suitable validity period, payment terms and supporting documents?Have the outstanding point checked.

Review the exported document as well as the data entry screen. An option may appear to be included; a description may contradict the line item. Check the information required for your business against an up-to-date source: this checklist is not a legal review.

Good to know. On 18 July 2024, the CNIL, France's data protection authority, pointed out that AI results can seem plausible while being inaccurate. It recommends defining permitted uses and checking whether the provider may reuse the data. For testing, use a fictional job file with no information that could identify a client. CNIL recommendations.

A correct total does not prove that the right work has been priced. Software can add up the cost of an item perfectly even when it has no place in the job.

Let's plan how your estimates are prepared.

Tell me about your software, the information you enter and the checks you want to keep. Let's see what can be prepared automatically.

5.0/5 on Google · 16 reviews
Let's talk

What should block sending after the review?

Block sending if required information is missing, the document has changed or the recipient is still uncertain. Approval must apply to a specific version: the reviewed file, its attachments and the recipient's address. Any correction after approval must cancel that approval in the tool and trigger a new check.

I would ask for a final preview before approving. It should show the client's name, the full email address and the files to be sent. "Job file approved" becomes too vague as soon as there are several revisions.

Blank approval worksheet to complete each time you send an estimate

CheckDetails to enterApproval requirement
Exact documentReference and version: ____The exported file has been reviewed.
Client and recipientIdentity and email: ____The address matches the job file.
Outstanding pointsChecks still needed: ____No essential check remains unresolved.
Message and attachmentsItems checked: ____The content matches this version.
ApprovalPerson and date: ____Explicit approval recorded.
Subsequent changeChange identified: ____A new check before sending.

Approval applies to one version, not to the job file forever. Changing a quantity, a term, an attachment or the email address must return the document to review.

Keep the statuses "approved by you", "sent" and "receipt confirmed" separate. Approval does not prove that a document was sent. A sending record alone does not prove that the client received the document, still less that they accept it.


What tests should you ask for before using this process?

Ask for a demonstration that deliberately triggers possible errors, using fictional job files and a test inbox you control. Write down the expected result before each action. A tool that produces an attractive document must also show that it can stop sending, retain the correct status and report a failure without announcing success.

  1. Missing information: remove a quantity. Expected result: a request for review and nothing sent.
  2. Changed version: approve the file, then change a work item. Expected result: the previous approval is canceled and new approval is required.
  3. Approved but not received: simulate a sending failure. Expected result: a visible failure, the document retained and no indication of receipt.
  4. Wrong recipient: replace the address with a different test address after approval. Expected result: sending stops, with a new preview and new approval required.

Do not run these tests using real clients' addresses. For each case, keep the expected result, the observed result and any correction, then repeat the test after making the correction.

Measure the work through to the checked document, not just the first draft. For comparable jobs, record the time spent preparing, correcting, reviewing and sending, along with any rework needed. Compare this with your usual approach. Without that record, I would not promise any time savings.

When commissioning this process, include these tests in your request for an automation quote. The rules for following up after sending can then be addressed separately.


What else should you ask before preparing the first document?

The choice depends on the information available and the checks you can actually perform. Existing software, an old estimate or a photo can help, each with different limits. Clarify the following points before giving them a role in preparation, without assuming a level of reliability they have not demonstrated.

Not necessarily. Check whether your software can accept the prepared descriptions, retain your prices and require approval before sending. Copying manually may be enough for an initial test. Ask for a demonstration of the connection between tools before choosing it.

Yes, as a starting point to check. An old description may be useful even when its price or terms no longer fit. Find the current source. For tests, remove information that could identify clients and check how the service uses the data.

A photo may help describe a situation without confirming dimensions, constraints or the condition of the underlying surface or structure. These details still need checking when they are required to price the work. The tool must flag unknowns and distinguish a visual estimate from confirmed measurements.


What should you keep in front of you before sending?

Keep the source of every important piece of information, the exact approved document and its recipient in front of you. Also plan how to handle a change or a sending failure. You can then judge the tool by the document actually checked and its behavior when something is missing, rather than by writing speed alone.

Summary of checks before sending an estimate

QuestionKey pointPitfall to avoid
Where did the line item come from?A verified source.Plausible but invented information.
What did I approve?A specific document and recipient.Approval that remains valid after a change.
What happened?A status consistent with the available evidence.Confusing approved, sent and received.

If you are considering automation suited to your business, you can tell me how you prepare estimates. Your current software and the checks you want to keep will be the starting point.

Further reading: the France Num guide to estimating software, published on 17 January 2023 and updated on 4 September 2026, covers templates, price libraries and document tracking features.

Portrait of Wail Saidani, founder of DevDream
Published on 30 October 2026 · updated on 30 October 2026 Wail Saidani

Founder of DevDream, a web and AI workshop for contractors and small businesses in Villeneuve-la-Garenne, near Paris, France (92). I build websites and set up automation around the tools small businesses use.

Follow me on LinkedIn →

An estimate prepared. The decision stays yours.

Let's discuss how you price work and what needs checking before you send an estimate. I can help you define an automation project that fits your tools.

Discuss my project →