The number one thing I would fix now with AI if I were still a GC

Written by Sarah Irwin, former in-house General Counsel turned community builder for in-house legal teams globally at ITGC.

Since I left in-house life, the question I get asked most about AI at ITGC dinners, webinars and calls is some version of this: “if you were still a GC, what would you do differently with AI?”

My answer has evolved somewhat over the last year, in particular since running an in-depth survey this summer (The State of In-house Legal Innovation in Europe) - which 82 GCs, CLOs and senior in-house legal leaders across Europe completed on a range of issues like AI, legal tech, budgets and more.

When I read through the results, one data point jumped out: 70% name accuracy as their single biggest concern with AI, but not one single person selected “it will replace me”.

Combined with live conversations I have across our events at ITGC, I came to one conclusion: legal teams have absolutely moved past whether to use AI, and how to build their confidence with it. The real problem now is that the AI doesn’t know their business.

After running our ITGC survey, Summize published its AI Fluency Report, based on a survey of 505 in-house legal professionals in the US. I was delighted to see that it explains why accuracy matters so much, where most teams are getting stuck, and how to fix it. It got me thinking - accuracy is the number one thing I’d fix now if I were still a GC. Here’s why.

Accuracy depends heavily on context

An AI answer can be technically correct and still wrong for your organisation. Legal advice is shaped by your risk appetite, your contractual positions, your internal policies and the commercial priorities of the business at that moment. A general-purpose model knows none of that unless someone tells it.

Lara Trope, Implementation Expert at Summize, gives a great example in the report of the difference between a simple piece of information and context:

Information = Your standard payment term is 30 days.

Context =

• You can accept 60 days for certain strategic customers without escalation.
• Anything longer needs CFO approval.

The context captures the rule, the reason behind it, where the flexibility sits and what should happen when a deal falls outside it.

The research backs this up: recent work by Stanford University's RegLab found that improving how relevant legal information is retrieved and applied increased accuracy from 70% to 83% in a benchmark of AI statutory research. AI performance depends on more than the model. It depends on whether the AI can access and use the right information.

Put simply, context is accuracy.

Lawyers know context matters, so why is accuracy still the biggest concern?

It seems there is a huge (context) gap between (1) lawyers understanding deeply how important their organisational knowledge is to any business decision and (2) embedding that knowledge into AI systems.

This is the finding from Summize’s report that surprised me: 75% of respondents say their AI tools always or frequently reflect their organisation's policies, preferences or historical decisions. Yet only 14.5% have connected AI to approved internal knowledge sources.

So how is that context getting into AI systems?

According to the report:

• 34.5% use saved prompts, templates or playbooks
• 27.7% type the context in manually each time
• 14.9% reference internal documents, contracts or policies
• 8.5% don't provide internal knowledge to AI at all

Summize describes this as teaching AI one prompt at a time, and I recognise it immediately because I've done it myself. The quality of the answer then depends on who is prompting that day and what they remember to include. The knowledge stays in individual lawyers' heads, which is exactly where it was before AI arrived.

Why this matters for Legal's reputation

There's a bigger issue here for in-house teams. When AI-assisted work goes out the door, Legal owns its accuracy. If an answer is technically sound but ignores your agreed fallback position or your risk appetite, the business will hold Legal responsible, and that chips away at trust the function has worked hard to build.

The report shows that 96% of respondents are confident they can spot an inaccurate AI response, yet only 30% would first verify it against a primary source. I don't doubt most lawyers would catch an obvious error. The harder ones to catch are answers that look right but miss how your business actually operates.

For lean teams, checking every output line by line isn't realistic. That's why the fix needs to sit upstream, in the knowledge the AI draws on.

How I would fix this

If I were still a GC, I would treat organisational knowledge as part of the AI system rather than something people add to a prompt when they remember. Lara's advice in the report is to start small while keeping the bigger picture in mind, and that is where I would begin.

1. Pick one repeatable workflow. Choose something low-risk and highly repeatable, such as NDA review. You don't need to map everything before you start.
2. Capture the reasoning behind your positions. Giving AI your NDA template is a start. Giving it the thinking behind the template is what improves accuracy: what you are trying to achieve, your standard position, where you have flexibility, what is unacceptable and what would make you escalate.
3. Give that knowledge an owner. Outdated context produces outdated answers. Someone needs to be responsible for reviewing and updating it, and that review should sit inside processes you already run rather than becoming another one-off project.
4. Test it, then build outwards. Run the context through a real piece of work. Does the output reflect your position? Does it know when something needs human review? Once it works, ask whether the wider legal team could use it, and eventually the business.

The next step towards AI fluency

I speak five languages, so I love the concept of “AI fluency”, which Summize defines as applying AI within workflows, connecting it to trusted organisational knowledge, critically evaluating its outputs and knowing where human judgment must stay central.

Most in-house teams are already strong on confidence and governance, and of course, inherent understanding of the importance of context. Connecting AI to what the business knows is the step that turns individual skill into something the whole team can rely on, and it's the most direct answer to the accuracy concern our community raised.

Read the full AI Fluency Report from Summize for the complete findings, the four stages of AI fluency and more practical advice from their team by clicking below. https://www.summize.com/ai-fluency-report