Clinical documentation improvement: what CDI programmes change
Short answer: clinical documentation improvement is the practice of making the medical record accurate and complete enough to reflect how sick the patient actually was. CDI teams read charts, spot where the documentation does not support the clinical picture, and query the physician. It exists because reimbursement, quality scores and case-mix index are all computed from documentation rather than from care.
What is clinical documentation, and why it needs improving
Clinical documentation is the record of what a clinician observed, concluded and did. It serves three audiences simultaneously and serves them differently.
The next clinician needs the reasoning. The payer needs specificity sufficient to assign a code. The quality programme needs structured elements that may have nothing to do with the encounter’s clinical substance.
Physicians write for the first audience, which is correct. The gap between that note and what the second and third audiences require is the entire reason clinical documentation improvement exists as a discipline.
A concrete instance: a physician writes “sepsis” where the chart supports severe sepsis with organ dysfunction. Clinically, the treatment is identical and the note is not wrong. For coding, severity of illness and expected mortality, they are materially different records. A CDI specialist spots the gap and queries the physician to document what the labs already show.
What clinical documentation improvement programmes actually do
Four activities, in a loop.
Concurrent review. Reading charts while the patient is still admitted, when a query can still be answered from memory rather than reconstructed. Retrospective review works but yields lower response rates and worse answers.
Querying. Asking the physician to clarify or document something the record supports but does not state. Queries must be non-leading, which is a real constraint rather than a formality: a query that suggests the answer is a compliance problem and an audit finding.
Education. Feeding patterns back to clinicians so the same gap stops recurring. The highest-leverage activity in any CDI programme and consistently the first thing cut when review volume rises.
Measurement. Case-mix index, query response rates, denial rates, and the agreement between coded severity and clinical severity.
CDI clinical documentation improvement teams sit in different places on the org chart depending on the institution, and the phrase clinical documentation integrity has largely replaced the older wording, on the argument that the goal is an accurate record rather than a more lucrative one. Both names describe the same four activities.
Programmes badged clinical documentation excellence tend to be the ones that took the education activity seriously; those badged clinical documentation improvement CDI in the reporting line of the coding department tend not to have.
Where CDI programmes drift is in treating query volume as the output. It is not. It is the cost.
What is clinical documentation improvement worth measuring
Anyone asking what is clinical documentation improvement in budget terms should start here, because the answer most programmes give is the wrong one. Most CDI reporting counts reviews performed and queries sent. Both are activity measures and both improve when a programme gets less efficient.
Better: measure the share of queries that change a code. A programme sending half the queries and changing the same number of codes is a strictly better programme, because every unnecessary query spends physician attention, which is the scarcest resource in the building and the one CDI is most resented for consuming.
Track query response rate by department too. A department below the rest is either not receiving the queries, not understanding them, or has decided they do not matter. Each has a different fix and the number alone does not distinguish them.
Clinical documentation integrity software, and its limits
Most clinical documentation integrity software does the same core job: read the chart, apply rules or a model, and surface charts likely to have a documentation gap. It is a worklist prioritiser.
That is genuinely useful, because the alternative is reviewers working through admissions in whatever order the census produced. A good prioritiser puts the charts most likely to yield a code change at the top, and a CDI team’s productivity is close to linear in how well that ranking works.
Two limits worth knowing before you buy.
The rules encode a general model of where documentation gaps occur. Your gaps are specific to your case mix, your physician group’s habits and your payer mix. Expect meaningful tuning, and ask during evaluation whether tuning is configuration or a support ticket.
And precision matters far more than recall here. A system flagging 40% of charts has not prioritised anything. Ask vendors for the share of flagged charts that produced a query, measured at a comparable institution, rather than for a detection rate.
Where ambient capture changes the CDI workload
This is the part that has genuinely shifted, and it shifts the work rather than removing it.
Ambient systems capture the consultation and draft the note from the conversation. The clinical content is therefore closer to what was actually said, which removes a category of gap that arose from physicians writing notes hours later from memory.
What it does not remove is the specificity problem. A note drafted from conversation reflects how clinicians talk, and clinicians do not talk in coding language. Nobody says “severe sepsis with acute kidney injury” out loud; they say the patient is septic and the creatinine is up. The ambient note is more accurate and no more codeable.
The opportunity is that a system generating the note can be built to flag, at the point of drafting, where the transcript supports greater specificity than the draft states. That converts a retrospective query into a prompt while the clinician is still in the encounter, which is the difference between a two-day loop and a two-second one.
That is a design decision made when the note-generation pipeline is built rather than a feature added later, and it is one of the things we build into AI medical scribe development. The mechanics of drafting from a transcript are covered in AI clinical documentation and the note structure itself in what an AI medical scribe note looks like.
Denials are the feedback loop nobody wires up
The clearest signal about documentation quality arrives months later, in the denial letters, and in most health systems it never reaches the CDI team.
A payer denying a claim for insufficient clinical support is telling you precisely where the record failed, on a specific case, with a reason attached. That is better evidence than any prospective rule set, and it is free. Yet denial management usually sits in revenue cycle, CDI sits somewhere else, and the two exchange summary statistics rather than cases.
Wiring that loop is mostly an integration problem rather than a clinical one. Denials matched back to the charts that generated them, grouped by reason and by service line, reviewed monthly by the people who write the queries. Programmes that do this find their query targeting improves faster than any vendor rule update delivers.
The same applies in the other direction. Queries that were answered but produced no code change are a list of things your team should stop asking about, and almost nobody keeps that list.
What this costs a physician
Worth stating plainly, because it is the reason these programmes meet resistance.
Every query is an interruption. It arrives in an inbox that is already the subject of most burnout research in the field, asks about a patient seen days ago, and is written in language designed to be non-leading, which frequently makes it harder to parse than a direct question would be.
Two things reduce the friction without reducing the yield. The first is precision: fewer, better-targeted queries, which is the metric argument above restated from the clinician’s side. The second is placement: a query answered inside the chart, in context, at the moment the note is signed costs a fraction of what the same query costs two days later in an inbox.
That second one is an engineering decision. A note-generation pipeline that already holds the transcript, the draft and the structured data can identify a specificity gap before the clinician signs. Doing it afterwards means finding them again, and finding them again is the expensive part.
Building a CDI programme, or extending one
There is no shortage of clinical documentation improvement toolkit material from the professional bodies, and it is worth reading, but a toolkit describes the activities rather than the sequence. If you are starting, the sequence that works is unglamorous.
Begin with a baseline: pull three months of coded records and measure the agreement between coded severity and what the clinical detail supports. That number is your opportunity, and it is usually concentrated in three or four DRG families rather than spread evenly.
Scope the first phase to those families. A programme that tries to review everything reviews nothing well, and a narrow start produces a defensible number within a quarter.
Buy the worklist prioritiser rather than building it, unless your case mix is unusual enough that vendor rules genuinely do not fit. Build the feedback loop, because no product will tell your cardiology group what its specific habit is.
And decide early who owns physician education. CDI programmes that report into coding treat education as optional; those that report into the medical staff treat it as the point. The second kind works better, for reasons that have nothing to do with software.
Clinical documentation improvement companies and services
The market splits three ways. Consultancies run assessments and build programmes. Outsourced CDI services supply reviewers, usually offshore and usually priced per chart. Software vendors sell the prioritiser and the query workflow.
Outsourcing review is the option to examine most carefully. It works where the constraint is reviewer capacity and the case mix is conventional. It works poorly where the value depends on knowing the physicians, because a query from someone the clinician has met is answered at a materially higher rate than an identical query from a queue.
Clinical documentation services of any kind should be judged on the same metric as an internal programme: code changes per query sent, not charts reviewed.
The takeaway
Clinical documentation improvement exists because the record has to serve payers and quality programmes as well as clinicians, and physicians reasonably write for the clinician. Measure code changes per query rather than query volume, buy the prioritiser and build the feedback loop, and treat ambient capture as a chance to move the query into the encounter rather than as something that removes the need for one.
EpochC builds AI medical scribe development for health systems, including specificity prompting at the point of note generation. See the AI medical scribe case study, or start a project.