AI Marketing for a Construction Company: Case Studies, Bids and LinkedIn

Coding Liquids tutorial cover featuring Sagnik Bhattacharya for AI Marketing for a Construction Company: Case Studies, Bids and LinkedIn.
Coding Liquids tutorial cover featuring Sagnik Bhattacharya for AI Marketing for a Construction Company: Case Studies, Bids and LinkedIn.

Treat each finished project as one source with three outputs. Collect a fact sheet, a photo log and a short site-manager voice note at handover, get the client's written consent, then have AI draft a one-page case study, a bid-ready version of it and four to six LinkedIn posts. You supply every fact and approve every claim.

Commercial buyers (developers, facilities managers, procurement teams) look for proof before they invite a contractor to bid, and most contractors' proof is a gallery of photos with no numbers. The case study supplies the numbers, the bid sheet shows relevance to a particular tender, and the posts keep the firm in front of the same buyers between tenders.

Follow me on Instagram@sagnikteaches

The example project is invented: a 3,800 m² two-storey office and warehouse fit-out for a logistics tenant, 14 weeks on site, with the neighbouring unit occupied throughout. Names, values and figures are illustrative.

Connect on LinkedInSagnik Bhattacharya

The risks sit in two places: facts an assistant invents to fill gaps, and photos or names the client never agreed to publish. The workflow below puts a gate at each.

Subscribe on YouTube@sagnikteaches

Collect the facts at handover, while the site manager remembers them

A fact sheet takes about 90 minutes, and it is the only input the assistant is allowed to use. Fill it from records, not from memory: the programme, the site diary, the safety log and the valuation. The "source of truth" column matters most, because it tells the reviewer where to check each line.

FieldExample entry (illustrative)Source of truth
Project and typeOffice and warehouse fit-out, two floorsContract
Floor area3,800 m²Measured area schedule
Contract value band$2m to $3m (band only)Final account
ProgrammePlanned 12 weeks; client-instructed mezzanine change in week 5 added 2 weeks; handover in week 14Programme and instruction log
Main constraintNeighbouring unit occupied; noisy work limited to agreed hoursSite diary
What we did about itThree phases; loudest work in weeks 1 to 4; heavy deliveries before the neighbour's start timeMethod statement and diary
Result at handover23 snagging items; all closed within 10 working days; no complaints logged by the neighbourSnagging list and complaints log
Safety[From the safety log: hours worked, recordable incidents]Safety log
Team rolesProject manager, site manager, two subcontractor leadsProject directory
Client comment[Client's own words, approved by the client]Client email
Consent tierA, B or C (see below)Signed consent

Add a photo log: 12 to 20 images, each with a caption a person wrote (stage, area, what it shows, date). The captions matter because the assistant cannot work them out: OpenAI's image-input FAQ says ChatGPT does not read file names or metadata, so the date your phone stored never reaches the draft. Claude's help page allows up to 20 files per chat, which fits a 16-photo log with the fact sheet alongside. Convert photos your phone saved as HEIC to JPEG before uploading, and check each assistant's help page for the formats and size limits it accepts today. If your handover pack already holds the dates and snagging data, the handover pack checklist shows how to gather them.

Finish with a two-minute voice note from the site manager answering three questions: what was hardest, what did you change, and what would you do differently. Transcribe it with your phone's dictation. The transcript gives the case study the lessons that no spreadsheet holds. If your team already uses a photo app on site, the photo documentation app comparison explains what its AI features do and which plan includes them.

Ask for consent before the final account is agreed

Consent is easiest to get while the client is pleased and the relationship is warm, and hardest to get six months after a dispute. Ask when you agree the final account, and ask for one of three tiers.

TierWhat you may publishWhat stays out
A: namedClient name and logo, an approved quote, photos, value bandSecurity details, exact value, personal data
B: anonymised"A logistics operator", photos with no identifying features, scale in broad termsName, location, distinctive features
C: privateNothing public. The project can be named as a reference in a bid if the client agrees each timeEverything in public

Ask for it in writing. The email below is the kind of draft an assistant writes well, because it is polite structure around facts you provide.

Subject: Featuring [project] as a case study: your choice of three options

Hi [first name], thank you for a good project. We would like to feature it as a one-page case study of about 400 words with up to eight photos, using your company name and our contract value band rather than the exact figure. You would see the draft and approve every word and image before anything is published. If you would rather we leave your name out, or not publish at all, either is fine. Option A is named, option B is anonymised and option C is private. Reply with the letter you prefer. Thank you, [your name]

Put the principle into the contract as well. The clause below is wording to discuss with your lawyer, not wording to copy.

After completion the Contractor may photograph the completed Works and publish the photographs, the Client's name and a factual description of the Works in its marketing material. The Client may approve the final images and text before publication. The Contractor will not publish images showing security measures, confidential equipment, personal data or identifiable individuals without their consent.

Consent covers more than the client. Operatives in a photo should agree to appear, and subcontractors' logos and a neighbouring tenant's signage can identify people who never agreed to anything. The construction AI policy tutorial covers how to write these rules down for the whole company.

From fact sheet and photo captions to a one-page case study

Start a fresh chat for each project. If two projects share one conversation, details such as a mezzanine or a delivery time can leak from one into the other, and you will not notice. Then paste the fact sheet and captions into a prompt with firm rules.

Write a one-page case study (350 to 450 words) from the fact sheet and photo captions below.
Audience: procurement leads and facilities managers at logistics and industrial clients.
Rules:
- Use only facts in the fact sheet and captions. If something is missing, write [NEEDS FACT: what].
- Do not write a client quote. Write [CLIENT QUOTE, approved by client].
- State the programme exactly as given (planned, change, actual). Do not round it, and do not
  say early, on time or late.
- No safety statistics unless they appear in the fact sheet.
- Structure: Brief, Constraint, What we did, Result. Short paragraphs. No promotional adjectives.
- After the case study, list every number and claim you used, each with the fact-sheet row it came from.
[fact sheet]
[photo captions P01 to P16]

After one pass of edits the example case study reads as follows.

Fitting out a 3,800 m² logistics hub beside an occupied unit

Brief. A logistics tenant needed two floors of office and warehouse space fitted out for operations on a fixed date.

Constraint. The neighbouring unit stayed occupied throughout, so noisy work had to stay within agreed hours and deliveries could not block shared access.

What we did. We split the works into three phases and put the loudest tasks, partition framing and slab cutting, into weeks 1 to 4. Heavy deliveries moved to the early slot agreed with the neighbouring tenant. In week 5 the client instructed a change to the mezzanine, which added two weeks to the 12-week programme, and we re-sequenced the remaining phases around it.

Result. Handover took place in week 14. There were 23 snagging items, all closed within 10 working days, and the site diary logged no complaints from the neighbouring unit. [CLIENT QUOTE, approved by client]

Compare that with an illustrative first draft from an assistant given the same sheet but weaker rules. It said the project was "delivered ahead of schedule" (blurring a 12-week plan and a client change into one headline), added "zero incidents" (not in the fact sheet) and supplied a quote, "the team were a pleasure to work with". All three came out. Every number now traces to a fact-sheet row, and a person can check that in ten minutes. The claims list at the end of the prompt is what makes that possible.

The same facts, rebuilt as bid evidence

A bid does not want the case study. It wants evidence for the questions the evaluator will score. The fastest way in is a compliance matrix built from the tender documents: each criterion, its weight, the evidence you hold and the gap. Ask the assistant to extract it, then fill the "gap" column yourself.

Criterion (illustrative)WeightEvidence we holdGap and action
Three relevant projects, last 5 years, $2m or more30%Logistics fit-out ($2m to $3m); retail refurbishment ($3m to $4m)Need a third; ask the client of the warehouse extension for consent
Method for working beside occupied space20%Phasing from the logistics fit-out; method statement M-07Update M-07 for this site's hours
Programme and risk20%Programme template; the week-5 change as a lessonDraft the programme for this tender
Health and safety record15%Three-year statistics registerConfirm certificate expiry dates
Social value15%Apprentice hours registerAdd local supplier spend from the accounts

The weights add up to 100. Two cautions apply. Claude reads the text of PDFs up to 1,000 pages, but its help page puts a 100-page limit on reading visual content, so a matrix from a long tender pack comes from the text and you check the drawings yourself. And before you upload a tender pack, read its confidentiality terms; some clients forbid sharing the documents with third-party tools.

The prompt that builds the matrix is short, and its third line does most of the work.

From the tender documents below, list every evaluation criterion in a table:
Criterion (quoted word for word) | Weight (as stated) | Page and section | Evidence asked for.
If the documents give no weight for a criterion or sub-question, write NOT STATED. Do not estimate.
Do not merge criteria. List pass/fail requirements (insurance, certificates, forms) in a separate table.
[paste the instructions to tenderers and the evaluation section]

In an illustrative run on a tender that weighted only its main headings, a first draft without that rule split each heading's weight across its sub-questions, and the splits added up so neatly that they looked official. None of them was in the documents. With the NOT STATED rule the blanks showed, and the bid coordinator asked how the sub-questions were scored through the tender's clarification process. The separate pass/fail table matters as much: a missing certificate fails a bid however well the narrative scores.

For each project you cite, the assistant turns the case study into a past-performance sheet: project, client (with referee consent), value band, dates, key personnel and one paragraph of relevance written for this tender. That paragraph is the only part written fresh each time. Paste the criterion wording beside it and ask for relevance in the evaluator's own words, then check each phrase against the fact sheet and the word limit. Filled in for the logistics fit-out, the sheet looks like this (illustrative).

FieldEntry
ProjectOffice and warehouse fit-out, two floors, 3,800 m²
Client[Client name], a logistics tenant; consent tier A; referee consent on file, [date]
Value band$2m to $3m
Programme12 weeks planned; client change in week 5 added 2; handover in week 14
Key personnelProject manager [name] and site manager [name], both proposed for this tender
RelevancePhased works beside an occupied unit; agreed noise hours; a mid-project change re-sequenced with no complaints logged by the neighbour

The assistant's first relevance paragraph opened with "extensive experience of delivering fit-outs in live, occupied environments". One project is not extensive experience, and "live, occupied environment" overstated the job: the client's own unit was empty during the works, and it was the neighbouring unit that was occupied. The corrected paragraph says exactly that, one completed fit-out beside an occupied unit, and attaches method statement M-07 as the evidence. An evaluator can mark a method statement; an adjective gives them nothing to score. The tender writing tutorial shows how to draft the whole response and score it against the evaluation criteria.

What an assistant must never write into a bid

Some items in a bid carry legal or contractual weight, and they come from a register, never from an assistant.

ItemLives inOwnerCheck before every bid
Insurance certificates and cover levelsBid registerFinanceExpiry date and cover amount
Quality and safety certificationsBid registerSafety leadCertificate dates and scope
Safety statistics, last three yearsSafety logSafety leadFigures match the log
Turnover and financial statementsAccountsFinanceLatest filed figures
Referee names and consentCase-study sheetsBid coordinatorWritten consent on file
Key personnel CVsHR filesBid coordinatorPerson is available on the bid dates

The assistant may format these and may check that a number in the narrative matches the register. It may not supply them. An assistant that is allowed to fill a gap with something plausible will do it, and an invented referee or a lapsed certificate in a signed bid is a far bigger problem than a thin paragraph.

A LinkedIn month that fits a site manager's week

LinkedIn posts can hold up to 3,000 characters, according to LinkedIn's help page, which is about 450 words. The posts that work for a builder are far shorter, usually 120 to 200 words, and they come from named people. The example month has one post a week from the site manager, one from a director and one from the company page, and the company page's week-4 post is the case-study teaser.

WeekSite managerDirectorCompany page
1The week-5 client change and how we re-sequencedWhy we ask for photo consent at the final accountHandover announcement, with a link to the one-pager
2Working beside an occupied unit: our noise-hours ruleWhat an evaluator wants to see under relevant experienceApprentice spotlight (with consent)
3One thing we would do differently on the next fit-outHow we price a client-instructed changeA site photo with a one-line caption
4Snagging: how 23 items were closed in 10 working daysA lesson from a bid we lostCase-study teaser, with the client's approved quote

The routine is simple. Each person records a voice note on Friday, an assistant drafts the posts in their voice from three of their own past posts, and the person edits and publishes. Comments on clients' and consultants' posts are written by hand, because that is where the relationships are. The risks of handing the account over to automation are covered in the tutorial on what is allowed and what is risky, and the tutorial on keeping your own voice explains the style-sample method.

Voice note to post: the edit that matters

Here is a shortened, illustrative transcript from the site manager: "so week five the client moved the mezzanine, it added two weeks, we had to resequence, we got the cutting done before the neighbour's quiet period and held the deliveries to the early slot." An assistant with no rules drafts something like this.

Excited to share that our team successfully delivered a state-of-the-art logistics facility! Despite a client change, we remained agile and kept the project on track. A big thank you to everyone involved!

It says nothing, and "on track" hides the two weeks the change added. The site manager's edit puts the facts back.

Week 5 of the logistics fit-out: the client moved the mezzanine. It added two weeks. What we did: finished the slab cutting before the neighbour's quiet period, kept deliveries to the early slot, and re-sequenced the three phases around the change. The neighbouring unit logged no complaints. [Confirm against the site diary before posting.] The lesson I would pass on is to agree the programme effect of a change in writing the same day.

That version has numbers, a method and one opinion. It is also where the bracketed check earns its place: the post cannot go out until the site diary has been read.

The photo check before anything is published

An assistant cannot judge whether a photo is safe or allowed to publish. Someone who knows the site looks at every image, full size, against this list.

  1. Everyone visible is wearing the required protective equipment and working safely, with no unprotected edge in frame.
  2. Faces appear only where the person agreed.
  3. No other contractor's signage, vehicle registrations or neighbouring tenant's branding is legible.
  4. No whiteboards, drawings or screens with confidential information are in view.
  5. No security equipment, access points or alarm panels are visible.
  6. The client's tier allows this image.
  7. The caption matches the photo and the date.

The failure to expect is a hero photo chosen for the view, in which an operative stands near an unprotected edge. It is easy to miss in a thumbnail, and a client's safety manager will not miss it on the post. The check takes 20 minutes per project and costs far less than the conversation that follows a miss.

Time, cost and which plan to use

StageTimeWho
Fact sheet, photo log and voice note90 minutes per projectSite manager and project manager
Consent request and follow-up30 minutesContracts administrator
Case study draft with AI30 minutesOffice or marketing
Review of facts, claims and photos60 minutesProject manager
Bid-ready sheet for one tender30 minutes per bidBid coordinator
LinkedIn drafts and edits40 minutes a weekSite manager and director

The first four rows add up to 3.5 hours a project. For tools, an individual plan costs about $20 a month on either of the main assistants, but client photos and tender packs belong on a business plan. A ChatGPT Business seat lists at $25 a month, or $20 billed annually, and a Claude Team Standard seat at the same figures, both with a two-seat minimum, so two seats cost $50 a month or $40 a month on annual billing, which is $480 a year (list prices, October 2026). Anthropic's privacy centre states that by default it does not use inputs or outputs from its commercial products to train models (content you explicitly report as feedback is the exception), and OpenAI's pricing page promises no training on business data by default for ChatGPT Business, so confirm both on the current pages before you rely on them.

How to tell whether it paid off

Do not expect a number to move in a month. Track four things over six to twelve months: how many bids used a case-study sheet, how many of those bids were shortlisted compared with bids that did not, how often a buyer mentions a case study unprompted, and how long a bid pack takes to assemble. For LinkedIn, count named buyers (procurement leads, facilities managers) who connect or respond, not impressions. If the sheet makes bid packs faster but shortlisting does not change, the case studies are probably not close enough to the work you are bidding for, and that is a selection problem rather than a writing one.

Questions builders ask before publishing a project

Do I need the client's permission to show a finished building?

Often yes in practice. The contract, a confidentiality clause or the client's own policy may restrict publicity, and who owns the photos depends on who took them and what the contract says. Ask before you publish, ask separately for a named case study and for each social post, and keep the written reply on file.

Can the assistant write the client's quote?

No. Ask the client for a sentence in their own words and send them the final text to approve. An invented quote attributed to a real person is a false statement, even if the client would probably have agreed with it. Leave a clear placeholder in the draft so nobody publishes it by accident.

What if the project is recognisable even without the client's name?

Then an anonymised version still identifies the client. Leave out the sector detail, the scale and any distinctive feature, or keep that project to private bid packs only. When in doubt, ask the client which details they are happy to see in public.

Is a full case study worth it for a small job?

Use a lighter version: a fact sheet, five photos and one post. Save the full one-pager for projects that resemble the work you want more of, because those are the ones a bid evaluator will ask about. A short, accurate record of a small job still goes into your bid library.

Further reads

Sources: Claude's help page on file upload limits (updated July 2026); Anthropic's privacy centre on commercial data and training (page dated August 2026); Claude and ChatGPT pricing pages (October 2026); OpenAI's image-input FAQ, as summarised in October 2026; LinkedIn's help page on posting updates.

Want a case study routine your site team will actually use?

On a 1:1 call we'll look at how projects are recorded today, choose the handover fields worth capturing and decide which drafting steps can be handed to AI safely.

Book a 1:1 call with me