AI That Works Like a Research Teammate

You may already be using AI to draft emails, summarise documents or answer common customer questions. Those tasks are useful, but they still require you to provide the direction, check the output and decide what happens next.

The more interesting development is AI that can behave like a junior teammate: it receives a goal, investigates the problem, runs several steps, checks its findings and returns with evidence. That is the idea behind Faraday, an AI agent developed by London-based Inherent, a company founded by former Google DeepMind employees.

According to TechCrunch, Inherent says Faraday reproduced findings from published scientific papers while using a comparatively small model. The company reported that it outperformed larger models from Anthropic and OpenAI on that specific research task. For you, the important lesson is not the leaderboard. It is that a smaller, focused AI system may be more useful than a large general-purpose chatbot when the workflow is designed properly.

TL;DR

AI is moving from answering prompts to completing multi-step work and reporting how it reached an answer.

For a Malaysian SME, the practical opportunity is to apply this approach to repeatable tasks such as sales research, stock checks, customer follow-ups and document review—not to hand over important decisions blindly.

What This Means

Faraday was tested on replicating published scientific research. It had to work out which experiments were worth running, use available tools and compare its results with the original findings. Inherent also wanted the system to show what it calls “research taste”: the ability to choose sensible next steps instead of trying everything randomly.

This is different from asking a chatbot, “What should I do?” A chatbot usually responds to the information placed in front of it. An agent is designed to pursue an outcome. It may break a task into smaller steps, search for information, use software, evaluate results and continue until it has enough evidence to report back.

Inherent says Faraday runs on Qwen 3.6, a model with 27 billion parameters, rather than one of the much larger frontier models. The parameter figure and comparison come from TechCrunch. You should not read this as proof that every smaller AI model is better. The useful point is that performance depends on the complete system: the model, instructions, tools, feedback process and quality checks.

Inherent uses reinforcement learning, a method that rewards useful outcomes rather than merely teaching a fixed list of rules. In plain language, the AI receives feedback on whether its result was helpful and gradually improves its approach. It is similar to training a new staff member by reviewing not only the final answer, but also whether they asked the right questions and followed a sound process.

The practical lesson: a focused AI teammate does not need to know everything. It needs a clearly defined job, access to reliable information and a way to show you what it did.

How This Applies to Malaysian SMEs

1. Improve sales research before you make contact. If you run a B2B trading company, engineering supplier or professional service firm, your sales team may spend hours checking a prospect’s industry, locations, product range and likely needs. An AI agent could collect information from approved sources, organise it in your CRM and prepare a short account brief. It could flag unanswered questions for your salesperson instead of inventing details. Your team still decides whether the prospect is suitable and what message to send.

For a local example, a Johor-based parts distributor serving factories could ask an agent to review a list of existing accounts, identify companies with similar operating profiles and prepare suggested follow-up questions. The agent’s role would be research and preparation. The salesperson would verify the information, speak with the customer and handle the relationship.

2. Reduce delays in customer service. Many SMEs handle enquiries through WhatsApp, email, Facebook and phone calls. The problem is not always the number of messages; it is the repeated checking required before giving an accurate reply. An AI teammate could look up product availability, delivery rules, warranty terms or previous conversations, then draft a response for approval.

A Klang Valley retailer, for instance, could use an internal agent to check whether an item is available at a particular outlet, identify the relevant exchange policy and prepare a reply in English, Bahasa Malaysia or a preferred customer language. The staff member should approve sensitive replies, especially those involving complaints, refunds, personal information or promises about delivery.

3. Make stock and purchasing decisions more disciplined. Inventory problems often come from scattered information. Sales data may sit in one system, supplier updates in email and warehouse notes in spreadsheets. An agent could compare recent orders with stock levels, identify items approaching a reorder threshold and produce a purchasing review list.

For a food distributor in Penang or a hardware wholesaler in Selangor, the agent could highlight unusual demand, slow-moving items and supplier lead-time changes. It should not place orders automatically at the start. Instead, it can explain which records it used, show the assumptions and ask a manager to confirm the action. That creates a useful review step without removing human responsibility.

4. Turn standard operating procedures into daily support. Your business may already have procedures for opening a shop, checking a delivery, onboarding staff or handling a service call. The difficulty is keeping those procedures visible and consistently followed. A focused AI assistant could guide employees through the relevant checklist and identify missing information.

For a cleaning services company, the agent could prepare a job brief from the customer contract, list site requirements and remind the supervisor to upload before-and-after records. For an accounting practice, it could check whether a client submission contains all documents required for the next processing stage. The AI is not replacing professional judgement; it is helping your team avoid preventable omissions.

What a Useful AI Teammate Needs

Element	What it means for your business	Simple starting example
Clear outcome	Define what a completed task looks like	Prepare a verified customer follow-up brief
Approved information	Control which files and systems the AI can access	Use the current product catalogue and CRM only
Tool access	Let the AI perform limited actions when appropriate	Read stock records and draft, but not send, a reply
Evidence trail	Require sources, assumptions and action history	Show the document or record supporting each recommendation
Human approval	Keep a person responsible for important decisions	Manager approves refunds, orders and customer commitments

The table describes implementation principles rather than measured market statistics. The source article’s reported performance comparison applies specifically to research replication and should not be treated as a guarantee for sales, finance or operations work.

Practical Takeaways

• Start with one repeatable workflow that currently consumes staff time.
• Write the desired outcome in one sentence, such as “prepare a verified delivery-status reply.”
• List the systems and documents the AI may access, and exclude everything else.
• Require the AI to show its sources, assumptions and incomplete steps.
• Begin with draft-only actions before allowing any automated updates or messages.
• Measure whether the workflow reduces waiting, rework or missed follow-ups.
• Review errors weekly and improve the instructions, data and approval process.
• Keep customer, employee and supplier information protected through appropriate access controls.

The Bigger Picture

The long-term change is not simply that AI models are becoming larger or smaller. It is that business software is gradually becoming more able to perform connected tasks. A future system may read an enquiry, check the customer record, confirm stock, prepare a quotation and ask you for approval. That is a very different experience from opening a blank chat window and writing a prompt.

However, capability alone will not make an AI teammate useful. Your processes must be clear. Your records must be reasonably organised. Someone must own the workflow and check whether the result is correct. If your product names are inconsistent across spreadsheets, or if staff rely on undocumented exceptions, an agent will struggle just as a new employee would.

Inherent’s approach also offers a sensible business principle: use existing tools where they are strong rather than building every component yourself. The company reportedly had Faraday use OpenAI’s GPT-5.5 Codex for coding instead of creating its own coding tool, as described by TechCrunch. For your SME, that may mean connecting an AI assistant to the accounting, CRM, inventory or communication tools you already use.

Do not begin by asking, “How can we use AI everywhere?” Ask, “Which task has a clear outcome, repeatable steps and a sensible approval point?” That question will lead you towards a practical AI teammate—one that helps your people work with more consistency while leaving important business judgement in your hands.

Ready to Streamline Your Operations?

Your business should run itself. AutoRunBiz deploys AI agents to automate your daily operations — WhatsApp orders, invoicing, customer follow-ups, and accounting. Book a free 15-min ops audit to see where automation fits your business →