Generic AI acts like a new entry-level hire, permanently stuck at day one. The same can be said for AI that's been added to your existing tools, but not into your actual workflows. 

AI that is merely ‘bolted-on’ is limited in ways that a truly connected AI operating system is not. Their usefulness only goes as far as what that particular tool ’knows,’ and you're still left filling in the gaps and stitching the information together yourself. 

So unless your firm puts in a lot of ongoing time and effort training these types of AI, the output will not be work that follows your review standards, references the right historical data, or reflects how you talk to that particular client.

That's why so many accounting firms haven't yet moved past using AI for individual productivity gains to using AI in a way that scales to support firmwide success. It’s no secret that accountants are too busy and never have enough time or people. And now foundational AI is asking you to become a prompt engineer and a skill engineer too? 

This is where Karbon saw an opportunity to build on everything its practice management solution already does for accounting firms. When you use AI like Kai in Karbon, all of your context is already there: your client communications, your team's workload and assignments, your workflows planned out step by step. Karbon knows your firm and is already built like the manual on how to run it.

A reminder on the 5 pillars of AI success in accounting firms

We believe that realizing the value of AI to your firm requires an operating system with five pillars working together. These form the structure of AI at Karbon, underpinning Kai, AI agents in Karbon, the Karbon MCP, Karbon’s AI Notetaker, and more.

These five pillars will enable you to realize true AI ROI

The five pillars are:

1. Input. Complete, structured, connected data and feedback. The raw material for your AI.

2. Context. The skills, knowledge, and workflows specific to your firm and your clients.

3. Guardrails. A safe, secure environment with control over data, access, and actions.

4. QA. AI checking AI, humans in the loop, and alerts when anomalies appear.

5. Enablement. Best practices, training, and tooling that keep your people leading.

What goes in the manual

Context, the second of the five pillars of AI success, refers to the skills, knowledge, and workflows that are specific to your firm and your clients.

Context has two halves: 

1. Knowledge is what your AI knows: your clients, your standards, your history. 

2. Skills are what it can do: the processes it can actually run. 

Foundation AI models arrive with neither. It's precisely the combination that separates AI that adds firmwide value from a personal productivity chatbot, or an AI feature that only touches fragments of your workflows.

Building it out means three layers.

1. Foundational skills. The tasks common to every accounting firm. Summarizing client correspondence, drafting a status update, pulling together a document request. This is the baseline your AI should arrive with, not something you write yourself. With Karbon, these skills can be customized, but they come built in and run in the background. 

2. Firm knowledge. Your templates, checklists, service definitions, pricing structure, review standards, and policies. The material that makes an output recognizably yours.

3. Workflows. Workflows pull from skills and knowledge, so they are innately unique to your firm and clients. Your review methodology, how you scope a CAS engagement, the way your niche clients get handled differently from everyone else.

None of it works without part one. Clean, connected input is the material the manual gets written from, so Input (Pillar 1) isn't optional homework you can circle back to later.

With context, it's not just about volume

Here's where our view on context may differ from what you've heard. Most people assume context means giving AI more. More data, more history, more documents. But half of the context is the opposite. It's telling AI what to ignore.

Every useful manual does this. It tells you which steps to skip for this client type, which exceptions don't apply, which of the twelve documents in the folder is the one that matters. 

A model handed everything your firm knows will weigh a four-year-old email the same as last week's engagement letter. Context is what separates signal from noise, so the output is actually what and where you need it, rather than somewhere technically defensible but ultimately useless.

We saw this play out directly in our beta for Kai. Firms consistently ranked AI workflows as most useful when they drew on real historical numbers instead of standardized assumptions. Their unique context is what mattered. 

And what you gain from clean context isn't only better output. It's predictable output, which for a firm that signs its name to the work is the only kind worth having. Especially when trust is the number one reason small business owners choose their accountant.

That's also why prompting is the weakest form of context. A prompt is a page someone writes fresh each time and throws away. It dies with the session and lives in one person's head. When ten people rely on ten personal prompting styles, quality depends entirely on who happened to do the work. 

But write the standard down once, in the system, and quality becomes a property of your firm rather than a property of an individual. It's how you already think about review notes and workpaper standards.

This manual is also nothing like the binder on the shelf nobody opens. It gets used on every engagement, and it gets revised every time someone corrects an output. That's the difference between documentation and context.

This is why context compounds in a system like Karbon. A living library of firm knowledge and processes gets more valuable every month. There's a second payoff too. The same manual that informs your AI shortens onboarding, protects the firm when a senior person retires, and gives you a real answer to the question of what happens when your best reviewer takes a month off.

What this looks like in practice

Here are some of the same scenarios from part one, this time through the Context lens. These examples are how context translates across common accounting processes and deliverables.

The non-billable work

In part one, we looked at an agent redistributing a departing manager's client book. Input makes that data reachable, and Context is what makes the recommendations correct.

Your firm has rules about this that nobody has ever written down. This is institutional knowledge. Things like certain clients needing a manager with specific niche experience, some relationships that can't move without a partner conversation first, and staff you'd never put together on the same account. 

If you capture those rules in the manual, the agent can propose a contextually appropriate recommendation, rather than a mathematically even one that ignores everything that matters.

One firm in our beta program discovered exactly this when they tried to teach an agent their invoice review process. A threshold they'd been applying for years, capped at a specific percentage, had never been written down anywhere. It was something you just had to know, until now. With Karbon, you will tell Kai once and Kai will capture it, remember it, and action it when needed.

Client advisory

Part one covered assembling a client briefing based on the full client record including email threads, call and meeting notes, portal messages, documents, billing history, and every past and current project. Input has provided the client data, and now Context can supply the judgment.

AI anchored in both the engagement history and your firm's point of view knows the client restructured their entity last year, knows which colleague spoke with them most recently, and frames the recommendation the way your partners would. 

It surfaces the questions your firm always asks, because those questions have been written down rather than carried around in the minds of whoever happens to be in the meeting.

Service delivery

Reconciliations, transaction matching, and continuous ledger updates all depend on thresholds and standards that vary by firm, and just as often by clients. What counts as material for one client is noise for another. For example, one client's revenue recognition runs monthly while another's runs quarterly, and a construction client gets reviewed against WIP schedules that a services client doesn’t need.

Aider, Karbon’s solution for month-end close automation, applies the same close rules that are specific to each client no matter who runs it. This standardization is powerful and a big step forward for most firms. 

Still, what those rules cannot know is how a specific client operates: their materiality threshold, the vendor who always codes wrong, the seasonality that makes a particular month’s numbers look off, who on their team actually replies. That knowledge usually lives in one reviewer's head. But Kai can layer this information onto the client’s record in Karbon, so the reviewer inherits it during the period close.

3 questions to assess your Context readiness

1. Could a new hire and an AI agent learn your firm the same way?
   If a new manager asked how your firm runs a monthly close or reviews a return, is there something you'd hand them, or does the answer live in a few senior people's heads? AI can only be taught what has been written down or built into workflow templates.

2. Do your AI tools know your clients, or just the world?
   Test it this week. Ask whatever AI tool your team uses a question that requires engagement history or client-specific facts, something a second-year staff member could answer. If the response is generic, you have a context gap.

3. Can you train the AI you're using?
   Some tools let you build custom skills and load firm knowledge alongside a built-in library. Others are sealed boxes with a text field. If there's nowhere to put the manual, context can't compound. Every correction, every workaround, every hard-won exception disappears the moment the session ends, and next month you're teaching the same tool the same lesson all over again.

If you are not already using Karbon, this might seem like a daunting task. Start smaller than you think you should. Don't write the whole manual at once. Just pick one high-volume workflow, document how your best person does it, and make that your first chapter. Then build from there. 

Karbon was built so that context is there from the start:

• Kai, Karbon's built-in AI, arrives already knowing accounting, with a foundational skill library that expands continuously.

• A practice knowledge center that intakes your firm's SOPs and turns them into skills agents can run.

• Agentic workflows and workflow templates made unique to your firm, carrying your steps, standards, and assignments as structured data.

• Client and work records that give every skill something real to run against.

• Routines that let you schedule agentic workflows.

• Karbon MCP (in beta) and new ’bring your own data’ features in Karbon Analytics, so your firm's context reaches across your GL, tax, and payroll systems instead of stopping at the platform edge.

You've spent years building how your firm works. But the firms that scale past the founder's calendar are the ones that turn expertise into an asset, not a bottleneck. Context is that asset: the manual that finally captures and puts that knowledge to work at a scale that wasn’t possible before AI. 

Next in the series is Pillar 3: Guardrails (stay tuned).

This is the second of five articles. Part one covered Input, the connected data that gives AI something real to work from. If you haven't read it, start there.