AI research agent cost tracking is the practice of measuring and attributing the cost of each AI-powered research investigation — tokens, tool calls, reasoning compute, and retrieval operations — to specific clients, matters, and projects.
A single AI research investigation can cost £0.50 or £15+. The range is wider than any other agent type. A simple fact-check costs pennies. A multi-source due diligence investigation can burn through hundreds of thousands of tokens, trigger dozens of web searches, and require frontier reasoning models that cost 10–20x more than standard ones.
Most firms running research agents have no visibility into per-investigation costs. They see a monthly API bill and have no way to attribute it. This makes accurate client billing impossible and budget overruns inevitable.
Key Takeaway: Research agents are the most cost-variable AI tool. Track cost per investigation, not just monthly totals.
Why Are Research Agents So Expensive?
Research agents cost more than other agent types because of how they work. Every aspect of their operation multiplies token consumption.
Multi-Source Retrieval
A coding agent reads a codebase and writes code. A research agent reads dozens — sometimes hundreds — of sources. Each source must be fetched, parsed, and processed. A competitive analysis that reviews 30 company profiles, 15 industry reports, and 20 news articles consumes tokens at every step.
Each retrieval operation involves:
- Querying a search API or database (tool call cost)
- Fetching and parsing the result (input token cost)
- Extracting relevant information (reasoning cost)
- Storing findings in working memory (context window cost)
Multiply this across 50+ sources and the costs add up quickly.
Deep Context Windows
Research requires maintaining context across all sources. The agent needs to remember what it found in source 3 while reading source 47. This demands large context windows — often 100,000+ tokens — which are more expensive to process than short, focused interactions.
Reasoning Model Requirements
Simple agents can use lightweight models. Research agents often need frontier reasoning models to synthesise information from conflicting sources, identify patterns across datasets, and draw conclusions from incomplete data. These models cost 10–20x more per token than standard models.
A document summary on a standard model costs £0.05. The same summary on a frontier reasoning model costs £0.50–£1.00. Research tasks that require genuine reasoning cannot be downgraded without quality loss.
Iterative Search Loops
Research agents do not operate in a single pass. They read, analyse, identify knowledge gaps, search again, read more, and refine their findings. Each loop burns additional tokens. A thorough investigation might iterate three to five times before producing a final output.
Each iteration can double the running cost. An investigation estimated at £2 can reach £10 after five rounds of deepening.
High Tool Call Volume
Research agents call more external tools than any other agent type. Web searches, database queries, document retrievers, calculation tools, and citation checkers — a single investigation can trigger 20–50 tool calls. Each has its own cost, typically £0.01–£0.10 per call.
For the broader picture on why AI agent costs are hard to predict, see our guide on unpredictable AI agent consumption patterns.
What Do Research Agents Cost by Task Type?
Research costs vary by orders of magnitude depending on the task. The table below shows benchmarks for professional services workloads in 2026.
| Research Task Type | Sources Reviewed | Typical Token Range | Cost Range |
|---|---|---|---|
| Simple fact-finding (single source) | 1–3 | 2,000–10,000 | £0.10–£0.50 |
| Targeted question answering | 5–10 | 10,000–30,000 | £0.30–£1.50 |
| Multi-source literature review | 15–30 | 50,000–150,000 | £1.00–£5.00 |
| Competitive analysis (multiple firms) | 20–50 | 80,000–250,000 | £3.00–£10.00 |
| Deep investigative research | 30–100 | 150,000–500,000 | £5.00–£15.00+ |
| Due diligence (legal/financial) | 50–200+ | 200,000–1,000,000+ | £8.00–£20.00+ |
Why the Range Is So Wide
Three variables drive the cost difference between the cheapest and most expensive research tasks:
Scope: A “research this company” request is inherently cheaper than “research this market, its competitors, regulatory environment, and growth trajectory.” Scope is the biggest cost driver.
Model tier: The same research task on a lightweight model costs a fraction of what it costs on a frontier reasoning model. Simple fact-finding does not need frontier reasoning. Complex synthesis does.
Iteration depth: An investigation that finds the answer on the first pass costs far less than one that requires five rounds of deepening. Some topics have easily accessible information. Others require extensive digging.
How Do You Track Per-Investigation Costs?
Per-investigation tracking requires three things: a unique identifier, cost capture at every step, and attribution to a client or matter.
Assign a Unique Investigation ID
Every research request should receive a unique identifier at the point of initiation. All subsequent actions — API calls, tool invocations, token consumption — get tagged with this ID. Without it, costs from concurrent investigations become impossible to separate.
Capture All Cost Components
For each investigation, capture:
- Token costs: Input, output, and reasoning tokens consumed across all LLM calls
- Tool call costs: Web searches, database queries, document retrievals
- Compute costs: Time-based charges for hosted research environments
- Retrieval costs: Vector database queries, embedding generation, document parsing
Sum these for the total investigation cost.
Set Alert Thresholds
Research investigations can run away. A poorly scoped request can iterate indefinitely, consuming tokens at each loop. Set cost thresholds that pause or flag investigations when they exceed expected limits.
Practical thresholds:
- Warning at £5: Notify the requesting professional that the investigation is becoming expensive
- Pause at £10: Require human approval before the agent continues
- Hard stop at £20: Terminate the investigation and deliver partial results
These limits prevent a single runaway investigation from consuming an entire week’s budget.
Build a Cost Dashboard
Track these metrics over time:
- Average cost per investigation by type
- Cost distribution (median, 75th percentile, 90th percentile)
- Cost per client/matter
- Cost trend over time (are investigations getting cheaper or more expensive?)
- Percentage of investigations exceeding thresholds
For task-level cost measurement methodology, see our AI agent cost per task guide.
How Do You Control Research Scope to Control Costs?
The most effective way to reduce research agent costs is to control scope before the investigation starts.
Define Scope Upfront
Require the requesting professional to specify:
- What question needs answering (not “research this topic” but “answer this specific question”)
- How many sources are sufficient (3 is not 30)
- What depth of analysis is needed (summary vs detailed analysis vs full report)
- What output format is expected (bullet points vs structured memo vs full report)
A tightly scoped request — “Find the revenue growth rates for these five firms over the last three years” — costs a fraction of an open-ended one — “Research the competitive market.”
Use Tiered Models
Route research tasks to the appropriate model tier:
- Simple fact-finding: Lightweight model (£0.10–£0.50 per investigation)
- Multi-source review: Standard model (£1.00–£5.00 per investigation)
- Complex synthesis: Frontier reasoning model (£5.00–£15.00+ per investigation)
Most research requests do not need frontier models. Routing them correctly saves 80–90% on those tasks.
Implement Progressive Deepening
Start every investigation with a shallow pass. If the initial results are sufficient, stop. If they are not, deepen selectively.
This approach costs £0.50–£1.00 for the initial pass. Only investigations that genuinely need depth incur the full £5–£15 cost. In practice, 60–70% of research requests are answered adequately by the shallow pass.
The Quality-Cost Trade-Off
Not every research task needs the same quality level. An internal briefing note does not need the same rigour as a client-facing due diligence report.
Define quality tiers:
- Quick answer: Fastest, cheapest, single-source. Good for internal queries.
- Standard research: Multiple sources, basic synthesis. Good for most client work.
- Deep investigation: Exhaustive sources, frontier reasoning, iterative refinement. For critical client deliverables.
Matching quality tier to business need is the single biggest lever for controlling research agent costs.
How Do You Attribute Research Costs to Client Matters?
Attribution is what turns a monthly API bill into actionable billing data.
Tag at the Point of Request
When a professional initiates a research investigation, they should tag it with a client code, matter number, or project identifier. This must happen at the start — retroactive attribution is guesswork.
Build the tagging into the request workflow. If requesting research requires selecting a client and matter from a dropdown, attribution happens automatically.
Aggregate Into Billing Summaries
Roll up per-investigation costs into monthly client summaries. These summaries should show:
- Number of investigations per client
- Total research cost per client
- Average cost per investigation
- Breakdown by research type
This data feeds directly into client invoices — either as pass-through costs or as justification for billing rates that include AI research.
Handle Shared Research
Some investigations serve multiple clients. A market analysis might inform work for three different client engagements. Split the cost fairly:
- Equal split across benefiting clients
- Proportional split based on relevance or usage
- Charge to the originating client with the option to re-use results
Document your approach. Clients may question charges for research that also benefited other clients.
For the broader framework on AI agent cost tracking, see our cost tracking guide for professional services.
Frequently Asked Questions
Why are AI research agents more expensive than other agent types?
Research agents consume more resources than coding, drafting, or classification agents. They retrieve information from many sources (high token input), maintain large context windows (expensive memory), use frontier reasoning models for synthesis (10–20x cost premium), and iterate through multiple search-analyse-deepen loops. A single investigation can trigger 20–50 tool calls.
How much does an AI research investigation cost?
Costs range from £0.10 for simple fact-finding to £20+ for complex due diligence. The median for professional services sits at £1–£3. The variables are scope (how many sources), model tier (standard vs frontier reasoning), iteration depth (how many loops), and output complexity (bullet points vs full report).
How do you track AI research agent costs per investigation?
Assign a unique investigation ID at the point of request. Tag all subsequent API calls, tool invocations, and token consumption with that ID. Capture token costs, tool call costs, compute costs, and retrieval costs. Sum them for the total investigation cost. Attribute to a client or matter using the tag applied at request time.
How can you control AI research agent costs?
Four main levers: define scope upfront (specific questions, not open-ended topics), use tiered models (lightweight for simple lookups, frontier for complex synthesis), implement progressive deepening (start shallow, deepen only if needed), and set cost thresholds that pause expensive investigations for human approval.
How do you attribute research agent costs to client matters?
Tag each investigation with a client code or matter number at the point of request. Aggregate per-investigation costs into monthly client summaries. For shared research that serves multiple clients, split costs proportionally and document the methodology. Feed these summaries into your billing system.
Keito tracks AI research agent costs per investigation, attributed to clients, matters, and projects automatically. Start tracking research agent costs.