Turn ten years of invoices and quotes into an AI knowledge base
Your old invoices and quotes hold the answers to what things cost and who did the work. Here is a seven-step method to make them searchable and useful to AI.
Key takeaways
- Old documents are an asset that nobody can use until they are sorted, structured and searchable.
- Structure first, AI second. Jobs, suppliers, line items and prices in one database make every later feature easier.
- Short structured snippets let the AI answer from clean facts instead of re-reading raw files each time: cheaper, faster and easier to check.
- Keep it alive. Every new job adds data, so plan for quality checks, outliers, inflation and duplicates.
If your company has been trading for years, you are sitting on a record of what you sold, what it cost to deliver, which suppliers you used and how you priced it. It is probably spread across invoice PDFs, quote spreadsheets, project folders and email attachments. Everyone knows it is valuable. Almost nobody can use it.
This guide sets out a method for turning that archive into a knowledge base that an AI application can use. It applies to any company with years of documents: agencies, contractors, manufacturers, event and production companies, trading firms.
Why old invoices and quotes are an asset nobody can use
Ask a simple question such as “what did we charge for this kind of job in the last three years?” and the answer usually depends on whoever remembers. Documents are saved in different formats, named in different ways and kept on different people’s computers. Searching by file name finds the file, not the figure inside it.
The result is that every new quote starts from memory and a few old files, and nobody is sure whether the price is in line with what you have actually done before. The data exists. It is simply locked in a form that neither people nor software can query.
The method: seven steps from archive to AI knowledge base
Step 1: Inventory what you have and where
List every source: invoices, quotes, project records, supplier costings, treatments or storyboards, scope documents, emails with attachments. Note where each lives, in what format, and roughly how many years it covers. Include the awkward ones, like scanned paper and spreadsheets with merged cells. You are not fixing anything yet, only finding out what you own.
Step 2: Decide the questions you want answered
This step shapes everything after it. Write down the questions the business asks repeatedly:
- What did we charge or pay for this last time?
- Who did we use for this, and how did it go?
- What is the median price for this kind of job, and what is the range?
- Which jobs resembled this brief?
Questions tell you which fields to capture. If nobody needs supplier names, do not spend weeks extracting them.
Step 3: Sort and clean with the people who know the work
Software can read documents, but only your team knows that one quote was a rush job at a premium, another was a favour at cost, and a third was never approved. Sit down with the people who did the work and agree what counts, what to exclude and how to name things consistently.
Budget real time for this. In one of our projects, sorting the material with the team took two weeks, and it was worth it: everything built afterwards rests on that clean-up.
Step 4: Load everything into one database with structure
Move from folders to a database with a few clear entities:
- Jobs: one record per project, with client, date, type and outcome.
- Suppliers: one record per company or freelancer, spelled one way.
- Line items: each cost or charge on a job, linked to the job and the supplier.
- Prices: the figures themselves, with currency and date.
Keep the original document attached to each record, so any number can be traced back to its source.
Step 5: Extract short structured snippets
Rather than asking an AI to re-read a raw quote every time someone has a question, process each document once and store short, structured snippets, for example: job type, date, supplier, item, quantity, unit price, notes. These snippets become the working memory of your system.
Step 6: Build on top: search, a master price list, then generation
With snippets in place, features come in a natural order:
- Search. Ask in plain language and get the matching jobs and line items, with links to the originals.
- A master price list. For each item or job type, show the median, the average and the range across past jobs, rather than a single remembered number.
- Generation. Use the same data to draft something new, such as a proposal or budget from a set of requirements and a target budget, for a person to review and adjust.
Step 7: Keep it alive
A knowledge base that stops growing goes stale. Make loading a finished job part of closing it. Review the data regularly (see the list of things to watch below).
Why snippets beat re-reading documents
You could point an AI at a folder of PDFs and ask questions. It works for a demo and disappoints in daily use. Extracting snippets once is better for three reasons.
- Token cost. AI models charge by the amount of text they process. Re-reading long documents for every question multiplies that cost. A few lines of structured snippet cost a fraction.
- Speed. Less text to read means faster answers, which matters when someone is building a quote with a customer waiting.
- Traceability. A snippet says exactly what was extracted and from where. If a figure looks wrong, you fix one record and every future answer improves. With raw documents, you are hoping the AI read the right paragraph.
That is the approach behind the Bangkok Productions data platform: the AI reads a clean knowledge base, not ten years of loose files.
What you can analyse once it exists
Once jobs, suppliers and prices sit in one place, questions that used to be impossible become routine:
- How have prices for a given item moved across ten years?
- Which suppliers appear most often, and at what cost compared with others?
- Which types of job are most common, and which are priced inconsistently?
- How do our final budgets compare with what similar jobs cost in the past?
The first time a team sees years of its own work in one view, it often spots patterns nobody had time to notice.
What to watch after launch
- Data quality. Wrong entries produce confident wrong answers. Make errors easy to flag and fix.
- Outliers. One unusual job can distort an average. That is why the median and range matter alongside the average.
- Currency and inflation. A price from ten years ago is not a price for today. Show dates and adjust before relying on old figures.
- Duplicates. The same supplier spelled three ways, or the same job entered twice, quietly skews every total.
- Privacy. Supplier prices and customer details are sensitive; see our guide to using AI on company data privately and in line with PDPA.
Where to start
Pick one question your team asks often, then check whether your archive can answer it. If you are still deciding where AI fits at all, our guide to which business processes are worth automating with AI helps you rank the options. When you are ready to talk about a build, see our web and application projects or describe your archive and the questions you want answered, and we will tell you what looks feasible. For budgeting, read what a custom business app costs in Thailand.
Frequently asked questions
How many documents do we need before this is worthwhile?
There is no magic number. What matters is that the same kind of question keeps coming up, such as what a job cost last time, and that your archive covers enough similar work to answer it. A smaller, well-organised archive beats a huge messy one.
Can the AI just read all our old PDFs and emails directly?
It can, but results are slower, cost more per question and are harder to verify. Extracting short structured snippets once, and answering from those, is usually more accurate and easier to audit.
Who needs to be involved from our side?
The people who know the work: whoever did the jobs, priced them and dealt with the suppliers. Technology can read a document, but only your team can tell which figures were typical and which were exceptions.