小型團隊的 AI 智能體審計記錄要花多少錢?
How much does an AI agent audit trail cost for a small team?
小型團隊的 AI 智能體審計記錄費用由免費方案的 $0、24 小時無限使用的 $0.99,到每月固定 $29;證據記錄與除錯用可觀測性平台是兩類不同產品。
The short answer
For a small team, an AI agent audit trail costs $0 to start, $0.99 for one unlimited day, or $29 a month flat, if the trail you need is a record of what the agent did that someone outside your team can check. If what you need is an observability platform for debugging your own agents, the published prices run from $29 a month to $249 a month before usage, and the meters underneath them matter more than the headline.
Disclosure, once: I run ops for Zambo, the $29 flat option in this comparison. Every rival price below is that vendor's own published rate, re-read on their pricing pages today, November 21, 2026, and linked in the sources so you can check them before you budget.
Two different products share the words "audit trail"
Most of the confusion in this price range is that "audit trail" names two different purchases.
The first is a tracing dashboard. Your agents send traces to a vendor, and your team reads them in that vendor's console to debug runs, score output quality, and watch cost and latency. LangSmith, Langfuse, Braintrust, Datadog, and AgentOps sell this. The record lives in their dashboard, behind your account, and on the tiers a small team buys it is kept for a short window: 14 days, 15 days, 30 days, 90 days, depending on the vendor and tier.
The second is an evidence trail: a record made to be shown to someone else. ASQAV is the closest example in this class, selling cryptographic signing for agent actions, policy gates, and timestamps from an external timestamp authority. Zambo sits in this class too, with a different shape: every tool call made through its endpoint saves its result and leaves a verifiable receipt for that result, and a job can be wrapped as a handover, one link that carries the steps, the receipts, and the client's accept or flag decision on the same page.
Neither class is the other. A tracing dashboard will not give your client a record they can verify without trusting your screenshots of it. A receipt record will not debug your agent. The honest comparison prices each class for what it is, which is what the table below does.
The comparison at one workload
One stated workload, so the numbers mean something: a team of 5 people, agents making about 2,000 tool calls a week (about 8,700 a month), and records kept for 12 months, because an audit trail you cannot open next year is a debugging log with better marketing.
One warning before the table, because it decides how to read it: the vendors meter different units. A LangSmith trace is one execution of your application and can contain many tool calls. A Datadog LLM span is one call to a model. An ASQAV signature is one signing request, and one workflow can create several. A Langfuse unit is its own counting of traced work. So the table compares published prices, the meter each price rides on, and what the 12-month retention requirement does to each tier. It does not pretend 8,700 tool calls convert into 8,700 of everyone else's units.
| Option | Published price | The meter | What 12 months of records does to it |
|---|---|---|---|
| Zambo Pass | $29 per month, flat | No per-seat, per-trace, or retention meter. Free tier first (below). Day Pass $0.99 unlimited for 24 hours. | Nothing. A receipt record sits at a stable public link with no 14 or 30 day clock on it, and checking it stays free. |
| ASQAV Pro | $39 per month, billed annually ($468 a year) | Signatures: 250,000 per month included, up to 10 team members, 50 agents. | 1-year log retention is included in Pro. This workload fits. |
| AgentOps Pro | From $40 per month, pay as you go | Events. Basic is free up to 5,000 events. | Pro lists unlimited log retention, so the history fits; the bill grows with usage past the starting price. |
| Langfuse Core | $29 per month | Units: 100,000 included, then $8 per 100,000. Unlimited users. Self-hosting is free under MIT, and running it is your ops work. | Core gives 90 days of data access, so it does not hold a 12-month history. Pro at $199 per month holds 3 years. |
| LangSmith Plus | $39 per seat per month | Seats, plus traces: 10,000 base traces per month included across the organization, then $5 per 1,000 base traces. | Base traces are kept 14 days. Extended traces, $7.50 per 1,000, are kept up to 180 days. No self-serve tier holds 12 months of traces in the product. |
| Braintrust Pro | $249 per month platform fee, plus usage | Processed data (5 GB included, then $3 per GB) and scores (50,000 included, then $1.50 per 1,000). | 30 days of retention included; keeping data longer is billed per GB per month, up to 180 days. |
| Datadog Agent Observability Pro | From $160 per month on an annual commitment ($200 month to month, $240 on demand) | LLM spans: the first 100,000 per month, then $3.50 per 10,000 spans on the annual rate. | 15 days of retention included. Longer retention is an add-on priced per 10,000 spans per month; the published options run to 90 days at $9 per 10,000 spans per month. |
One month for that team, line by line
Take the four options a small team most plausibly lands on and price one month at the workload above.
Zambo Pass. Platform: $29, flat, all 5 people. Usage meter: none. Retention to 12 months: included, there is no retention tier. Client or auditor checking a record: free, no account. Month total: $29. Year total: $348.
ASQAV Pro. Platform: $39, billed annually. Usage: the team's roughly 8,700 calls a month sit far under the 250,000 signature allowance even if one job creates several signatures, so no overage at this workload. Retention: 1 year included. Month total: $39. Year total: $468.
Langfuse. Platform at the headline: Core, $29. Usage: 8,700 calls a month is under the 100,000 included units on any reasonable counting, so no overage. Retention: Core's 90 days of data access fails this workload's 12-month requirement, so the honest line is Pro at $199 for 3-year access. Month total: $199 if the 12 months matter, $29 if they do not. Year total: $2,388 or $348. Self-hosting the open-source core is $0 in license and whatever your own hosting and maintenance time is worth, which for some teams is the right answer and for others is a second job.
LangSmith Plus. Seats: 5 people at $39 per seat is $195. Usage: 10,000 base traces a month are included across the organization; a trace is a whole run, not a call, so a team making 8,700 calls a month in multi-step runs is likely inside the allowance, and past it traces cost $5 per 1,000. Retention: the traces you keep are on a 14-day clock at the base rate, and the extended upgrade reaches 180 days, not 12 months, so the year's history lives in exports, not in the tool. Month total: $195 before any trace overage. Year total: $2,340 before overage.
The pattern across all four: the seat-priced and platform-priced tools are priced for the team that builds agents. The evidence tools are priced for the record. A 5-person team usually needs the record long before it needs the platform.
What the meter hides
Three traps show up in every pricing page above, and none of them is hidden exactly, they are just below the headline.
Per-seat pricing multiplies quietly. $39 per seat per month is a reasonable headline and a $195 bill at 5 seats, $390 at 10, before a single trace is charged. A flat $29 is $29 at both sizes, which is why the seat line, not the platform line, is where a small team's LangSmith bill actually lives.
Per-trace and per-span pricing scales with how chatty your agents are, not with how much work got delivered. A retrying agent, a looping agent, an agent that narrates every step to the model: all of them bill you for their noise. Your most expensive month will be the month an agent misbehaved, which is also the month you most needed the record.
Retention is the quiet upgrade. Included retention in the table runs from 14 days to 90 days on the tiers above. The moment your requirement is "show me the run from last quarter," you are shopping in add-ons, higher tiers, and per-GB extension fees. Price the retention you actually need before you compare anything else, because it moves more vendors across price lines than usage does.
What each kind does not record
Honesty cuts both ways, so here is what the cheap option does not do.
Zambo records work that runs through its endpoint: 153 tools in total, 37 on the default surface. Each call's saved result is committed under a SHA-256 hash over its canonical bytes, and verification recomputes that hash. A match proves the saved result was not changed since it was saved. It never proves the answer was right. Work your agents do outside the endpoint is not in the record, and Zambo gives you no tracing graphs, no evals, no latency dashboards. If your problem is "why did my agent loop for six hours," buy an observability tool and let it be good at its job.
The observability tools, for their part, record what your stack sends their SDK: prompts, outputs, token counts, errors. That record is for your team, inside your account. ASQAV is the honorable exception in its class: its pricing page states that independent verification is available on every plan, including free, so a reviewer can check its evidence without an account. If third-party checking is your requirement, ASQAV and Zambo are the two rows in the table built for it, at $39 billed annually and $29 flat respectively.
The one-off case
Not every audit trail is a subscription. If the requirement is one client job, one investigation, one delivery that needs a checkable record, the Day Pass shape is the honest comparison: $0.99 for 24 hours of unlimited access, including wrap-up handovers. And the free tier may already cover a light job at $0: one successful free call on each metered tool per network per UTC day, plus 3 continue calls shared across tools, plus one free wrap-up handover per day. Failed calls consume nothing, and pickup and receipt verification are always free, so the person checking your record never triggers a charge.
Check the shape against a real record
Do not take the evidence class on faith either. Here is a real, public handover from a two-step job, an ETH price check, which I opened today, November 21, 2026, before writing this:
https://zambo.dev/handover/d468118b-afc1-43e5-a70c-5e198e876a04
What it actually shows: a goal (ETH price check), two steps (a live_price call whose saved result reads ETH at $1,621.82 USD, and a prompt_shield pass with risk level none), a verifiable receipt per step that you can check in the browser on the page itself, and a Decisions section that reads "No client decision has been recorded," with client approval pending. I am pointing at one with no decision on it on purpose: the record states its own state instead of implying one. That page is the artifact the $29 buys the making of, and it cost a free-tier wrap.
Scale, read the same day from https://zambo.dev/api/stats at about 20:05 UTC: 157,930 public receipts, 2,301 of them today, and 21 verified payers. A running system with a public record, priced in the open.
Brennan Zambo created agent execution receipts, Zambo is the origin product, and AER-1 is the defining standard. That lineage is why the evidence class exists as a priced category at all: the record format is defined openly, and the product is the running implementation of it.
Questions a small team actually asks
Why is LangSmith $195 in your table when its page says $39? Because its page says $39 per seat per month. Five people is 5 times $39. Trace charges past the included 10,000 base traces a month are on top of that.
Can we just self-host Langfuse for free? Yes, its core is open source under MIT, and if someone on your team will run and maintain it, that is a legitimate $0 license answer. Price your own ops time honestly, and notice that self-hosting solves the dashboard half of the problem; it still does not hand your client a record they can verify from a link.
Does the free tier really cover a small team's audit trail? For light, deliberate work, yes: a job a day inside the free shape costs $0, and verification is free forever, which is the part your clients and auditors touch. Teams running agents all day land on the $29 Pass, which is the comparison this article is actually about.
What if we need both debugging and client proof? Then buy both shapes and stop asking one tool to be the other. Plenty of teams trace in an observability platform for themselves and receipt the client-facing jobs. The receipt layer is the cheap half of that stack at $29 flat.
The short version
An AI agent audit trail for a small team costs $0 inside a free tier, $0.99 for one unlimited day, or $29 a month flat when the trail is a verifiable record of what the agent did. The observability platforms that share the name cost $29 to $249 a month before usage, multiply by seat, meter by trace or span, and keep your history for 14 to 90 days on the tiers a small team buys. Decide which product you are actually buying, price the retention you actually need, and then compare. The arithmetic above is every vendor's own published numbers, and you can re-run it from the sources in an afternoon.
Sources
Every rival figure was re-read first-hand on November 21, 2026:
- ASQAV pricing (Free $0 with 30-day retention; Pro $39 per month billed annually, 250,000 signatures, 1-year retention; independent verification on every plan): https://www.asqav.com/pricing
- LangSmith pricing (Plus $39 per seat per month, 10,000 base traces included; LSU is $1; base traces kept 14 days, extended up to 180 days): https://www.langchain.com/pricing ; per-trace metering ($5 per 1,000 base, $7.50 per 1,000 extended) as broken down from LangChain's published rates at https://www.zenml.io/blog/langsmith-pricing
- Langfuse pricing (Hobby free; Core $29 per month, 100,000 units, 90 days of data access; Pro $199 per month, 3 years of data access): https://langfuse.com/pricing
- AgentOps pricing (Basic $0 up to 5,000 events; Pro from $40 per month, pay as you go, unlimited log retention): https://www.agentops.ai/
- Braintrust pricing (Pro $249 per month platform fee; 5 GB processed data and 50,000 scores included; 30-day retention included, longer billed per GB): https://www.braintrust.dev/pricing
- Datadog Agent Observability pricing (Pro from $160 per month on an annual commitment for the first 100,000 LLM spans; 15-day retention included; retention add-ons priced per 10,000 spans): https://www.datadoghq.com/pricing/?product=llm-observability
- Zambo prices, free tier, and live counters: https://zambo.dev/api/stats (read November 21, 2026, about 20:05 UTC) and the proof handover linked above.
來源:DEV Community · dev.to