Key AI prompting techniques for finance, plus a ready-to-use skill to try

Written byAnna Katharina Bollé

AI & IntelligenceJuly 23, 202610 min read

Learn how prompt engineering for finance helps teams standardise AI instructions, organise unstructured data, and generate reliable, audit-ready results.

Key takeaways

• Structured instructions help you get reliable financial data you can easily explain to auditors.
• Pre-tested templates save you from spending time and money fixing errors.
• By structuring how AI handles unstructured data, your team can spend less time on manual entry and more time on strategy.
• No matter how advanced the technology, human review still matters for compliance.

Prompt engineering for finance is the process of structuring instructions, roles, data boundaries, and rules so generative AI models produce accurate and audit-ready results. That matters because AI can still miss context, hallucinate, or return outputs that are not reliable enough for finance work. In this article, we use the example of messy SaaS and AI spend to show how we tested practical techniques that make results more stable and trustworthy. We also share the finished skill so you can try it yourself, test it in practice, and get familiar with how prompting works.

What is prompt engineering for finance?

Prompt engineering for finance is the practice of setting up AI with the right instructions, roles, context, and source data so it can produce outputs that are accurate, repeatable, and easy to explain. In finance, that matters because the result needs to be accurate, useful, and easy for people to trust and review.

That is becoming more important as AI adoption keeps growing across UK finance. As these tools become more common, the way they are instructed matters more too. Financial data is sensitive, and standard chatbots are built to predict language, not handle exact figures. Structured prompting helps turn them into something more useful for daily finance work.

Why generic prompts fail in corporate finance

Many professionals think that using AI is as simple as asking a question. But when you use a generic prompt like “summarise this balance sheet,” the AI has to guess the context, starts to hallucinate, and the result often lacks the reliability finance work needs. In corporate finance, generic prompts often fail because they do not give the model enough structure.

A standard model might miss transaction lines, group unrelated items together, or mistake a vendor name hidden inside messy card text. It can also return an answer that sounds confident even when parts of the data are incomplete, or when the model just got lazy and stopped after half of it. In finance, that matters even more because real money is at stake and clear human accountability for technology pipelines is still needed rather than treating the algorithm as the accountable party.

Building a finance prompt to fix SaaS and AI spend oversight

We wanted to see how far prompting calibration could actually take us in a chat window when the problem was real, messy, and finance-shaped. So we took the real finance problem of getting oversight over SaaS and AI spend, which we had first seen in our own company. These findings were then confirmed even more clearly in our European SaaS and AI Spending Report 2026, where Moss analyzed €302 million in company card spend across Europe. The report showed how enterprise software spend keeps rising, and how that is being accelerated even further by fast AI adoption at the moment, with spending growing by 309%.

That creates a problem for finance teams. One team signs up for the hot new tool. Another team finds a new favourite a few months later. Someone adds a third because the free trial was too good to pass up. Give it a year or two and the company tool stack becomes a junk drawer, full of tools nobody quite remembers approving, some still billing every month, some nobody has opened since March.

And finance gets the fun part: finding them in the books. They rarely sit neatly where you would expect them. Some tools are booked to software. Some land in general admin. Some hide in prepaid expenses. Some show up once a year. Some quietly release from an accrual with no fresh charge to notice. Some arrive as vague card text that tells you everything except what was actually bought.

So when finance gets asked, “Can you quickly tell us what tools we are paying for?”, the answer is usually sitting somewhere inside thousands of transaction lines. Which is a very finance way of saying: good luck. And because we did not want to scan our own 15,000 Excel lines of transactions from last year by hand, we decided to test how long it would take us to arrive at a prompt that could do this task in a way that was both confident and trustworthy. It took a few iterations, and we learned a few things along the way, so let’s walk through them.

Key practical insights while optimising a finance prompt

To tackle this, we built a Software Spend Surface skill to find hidden SaaS spend from a raw transaction export. The skill processes the data and returns an alphabetical table of vendors, descriptions, and flags for review. The key learnings we had along the way:

1. Without enough instruction, AI sounds confident but misses things

When we first tested this task with a basic prompt, the AI found only 68 vendors against a manually checked baseline of 267. The answer looked usable and confident, which made it dangerous because it had missed nearly three-quarters of the data.

That is the problem with finance data. Without clear instructions, AI tends to give you the impression that it has completed the task properly. It will still sound helpful, even when it has only covered part of the file or left gaps behind. Finance exports are not one clean paragraph of context. They are rows, columns, booking texts, weird vendor names, duplicates, dates, currencies, and noise.

2. The skill needs a solid reference point

For finance data, the model cannot just “think about the file”. It has to run the work. When the task gets this big, that means scripts: small pieces of code that read the uploaded file, go through the rows, apply the same checks again and again, and return a structured result.

We could not just ask AI to look at thousands of transaction lines and magically know every software vendor. That gets expensive, inconsistent, and weird very fast. So the way we solved that was by backing the skill with a vendor taxonomy as a reference file. It gave the model a structured list of known SaaS and AI vendor patterns to match against, instead of asking it to guess from scratch.

Because we work with enterprise finance data, we could draw on real vendor patterns companies actually have in their books. That taxonomy helped with two things at once: finding better candidates and cleaning up messy names. Instead of letting the model decide whether “WWW.SPRINGER.COM”, “Springer Nature”, and some strange booking text belong together, the workflow had a known reference point.

3. The AI system changes the result

Once the prompt and the reference layer were in place, we realised something else: the system you run it in still changes the outcome. The same prompt or skill does not behave equally well in every environment.

We use different AI tools internally because they are better at different kinds of work. Claude handled this setup well because it could run the scripts in chat. Paid ChatGPT can work for the same reason. Gemini chat did not support the simple flow we needed. So the finding is not just “write a better prompt”. It is also that when you move finance workflows between AI tools, you move between different file handling, script execution, context limits, and model behaviour. The workflow you build in one system may not carry over cleanly to another, even if the instruction is identical.

4. Calibration costs money before it saves time

Iterating on a large Excel file is not free. You run the workflow, inspect the output, adjust the logic, add vendor patterns, run it again, compare the result, find a new failure mode, and repeat.

By the time we had tested this properly on a large file with many rows and columns, we had burned around $500 in token spend. Slightly rude. Also useful. It shows why a calibrated skill has value: someone already paid for the part where the model spends money being confidently incomplete.

5. AI is still probability

Even after engineering, scripts, taxonomy, and repeated testing, AI is still probability. Better probability, not certainty. We went from 68 vendors in the first broken run to around 97% coverage in the final test setup. That is a serious improvement. It is much closer to finance-ready than throwing a transaction list into an empty chat window and hoping the model behaves.

It is also much faster than asking a human to find every hidden SaaS vendor manually. But it still needs review. The point is not that AI replaces finance judgment. The point is that the first pass becomes good enough for finance to inspect, challenge, and improve.

How to test and run the software spend skill in your environment

If you want to see how the prompt works and which SaaS and AI vendors show up in your transactions, you can try this skill inside your company’s approved AI environment. That way it keeps your confidential data secure.

1. Download the skill file and save it locally on your computer.
2. Open the AI tool that your company’s IT and security policies allow you to use.
3. Add the skill file to that tool to load the engineered prompting instructions.
4. Upload a transaction export from your accounting system, such as DATEV, Xero, QuickBooks, or NetSuite.
5. Use a full-year export where possible, as spend often hides across different months.
6. Let the skill run its first pass, responding to any short clarifying questions if your data is unclear.
7. Review the resulting vendor table and check the rows marked with uncertainty flags.
8. Export the final table as an Excel or CSV file ready for your team’s review.

Choosing the right generative AI tools to test the prompt

To run advanced prompting skills successfully, you need the right generative AI tools. Not all platforms handle structured scripts and files in the same way:

• Claude: Works well in both free and paid versions. It can run the backend scripts that complex financial prompts need, making it a highly reliable option for this workflow.
• ChatGPT: Requires a paid plan to run the script-execution environment. The free version does not support these features.
• Other approved AI tools: Success depends on the specific setup. The tool must be able to read the full ledger file and execute scripts. If it cannot do both, it may return a complete-looking table that actually misses large chunks of your data.

Always remember that file sizes can affect stability. If accuracy is critical, run the skill twice in separate, clean chat windows and compare the tables. While matching outputs build confidence, they still do not replace the need for careful human review.

AI that frees up finance teams for higher-value work

Good prompt engineering for finance does not make AI the ultimate decision-maker. Instead, it makes AI an incredibly powerful assistant that does the heavy lifting before the final review: reading raw transaction ledgers, compiling clean tables, flagging unusual items, and highlighting uncertainty.

In audit and compliance work, technology cannot replace human judgment. The Financial Reporting Council explicitly warns that generative and AI agents do not dilute professional accountability. It is the humans, the firms, and responsible individuals who remain strictly accountable for audit quality and professional skepticism. By using calibrated prompting skills, your finance team can spend less time on manual data parsing and more time on strategic validation.